DECOMPOSITION APPROACHES FOR ENTERPRISE-WIDE OPTIMIZATION IN PROCESS INDUSTRY by NIKISHA K. SHAH A Dissertation submitted to the Graduate School-New Brunswick Rutgers, The State University of New Jersey In partial fulfillment of the requirements For the degree of Doctor of Philosophy Graduate Program in Chemical and Biochemical Engineering Written under the direction of Marianthi G. Ierapetritou, PhD And approved by ________________________ ________________________ ________________________ ________________________ New Brunswick, New Jersey May, 2015 ABSTRACT OF THE DISSERTATION Decomposition Approaches For Enterprise-Wide Optimization In Process Industry By NIKISHA K. SHAH Dissertation Director: Marianthi G. Ierapetritou, PhD Enterprise-wide optimization (EWO) has gain lot of interest in recent years as the globalization trends of past few decades have significantly increased the scale and complexity of modern process industry and increasing economic pressures to remain competitive in global marketplace. EWO entails optimization of supply, manufacturing, and distribution activities to reduce costs and inventories through an integrated and coordinated decision-making among various functions in the industry (vendors, production facilities, and distribution). One of the major challenges in achieving EWO is mathematical tools for planning and scheduling for manufacturing facilities. Main objective of this dissertation is to develop mathematical methodologies to assist in achieving EWO goals for chemical process industry. Specially, mathematical formulation for planning and scheduling decisions and decomposition strategies will be developed in order to bridge the gap between concepts and industrial application. In this work, planning and scheduling of multisite, multiproduct batch production and distribution faculties is addressed via dual decomposition based approach, which aims to reduce computational complexity through parallel computation. In the area of continuous production facilities, lack of efficient scheduling models prevents us from developing coordinated planning and scheduling tools. To address this issue, ii mathematical formulations for scheduling of refinery operations is developed and novel heuristics and mathematical decomposition strategies for large scale complex mixed integer linear programming models are proposed. Throughout this dissertation, case studies will be used to demonstrate the applicability of proposed decomposition approaches. iii Acknowledgements I would never have been able to finish my PhD without the support and encouragement from my family and my advisor. I have been extremely fortunate to have a wonderful parents, advisor, and friends supporting me for the last few years. Without all of your guidance and help, I would not have reached the finished line. I would like to express my deepest and sincere gratitude to my advisor, Dr. Marianthi Ierapetritou, for her excellent guidance, caring, tremendous patience, encouragement, and providing me with an excellent atmosphere for doing research and grow as an independent thinker. Her mentorship was paramount in helping me grow academically, professionally and personally. I am not sure many graduate students are given the opportunity to develop their own individuality or allowed to work with such independence. Thank you for always supporting me, challenging me to do my best, and believing in me when I had lost faith in myself. For everything you‟ve have done for me, Dr. Ierapetritou, I thank you. I would also like to thank the members of my committee, Dr. Ioannis P. Androulakis, Dr. Rohit Ramachandran, and Dr. Beverly Smith for their support and kind accommodations. I was fortunate to have Dr. Zhenya Jia introduce me to my research topic when I was still an undergraduate. I would especially like to thank Dr. Georgios Saharidis for his assistance and guidance in getting my interest in decomposition strategies started on the right foot and providing me with honest critical feedback on my first research article which has helped greatly improved my ability to communicate my research effectively. I wish to acknowledge past and present members of in the Laboratory for Optimization and Systems Analysis and friends at Rutgers whose friendship and support I will always treasure: Dr. Edgar Davis, Dr. Patricia Portillo, Dr. Hong Yang, Dr. Zukui Li, Dr. Vidya Iyer, Dr. Mehmet Orman, Dr. Aditya Vanarase, Dr. Fani Boukouvala, Dr. Yijie Gao, Dr. Kaiyuan He, Dr. Shuliang iv Zhang, Dr. Ravendra Singh, Dr. Amanda Rogers, Dr. Jinjun Zhuge, Nihar Sahay, Zhaojia Lin, Parham Farzan, and Sebastian Escotet. Lastly, I would like to thank my family for all their love, encouragement, and unconditional support. To my parents who raised me with a love of science, engineering, and critical thinking. Without inspiring example of my mother, her friendship, and her constant belief in me that I can achieve anything through hard work and dedication, I would never have made it this far. To my sister Khushbu, for teaching me the virtues of personal strength and perseverance – in the face of difficulties. To my younger brothers Vaibhav and Dhruv, for always providing me with an entertainment. v Dedication To my mother: Shobhana Shah vi Table of Contents ABSTRACT OF THE DISSERTATION ........................................................................................ ii Acknowledgements ......................................................................................................................... iv Dedication ....................................................................................................................................... vi List of Tables ................................................................................................................................. xii List of Illustrations ......................................................................................................................... xv Chapter 1 .......................................................................................................................................... 1 1. Introduction .......................................................................................................................... 1 1.1. Enterprise Wide Optimization ......................................................................................... 1 1.1.1. Problems and Challenges ......................................................................................... 3 1.2. Planning and Scheduling in Process Industry .................................................................. 5 1.3. Refinery Operations Scheduling ...................................................................................... 6 1.4. Motivation: development and implementation of decomposition tools for planning and scheduling in process industry ..................................................................................................... 7 1.5. Outline of dissertation ...................................................................................................... 7 Chapter 2 .......................................................................................................................................... 9 2. Planning and scheduling of multisite batch production and distribution facilities .............. 9 2.1. Introduction ...................................................................................................................... 9 2.2. Problem Statement ......................................................................................................... 15 2.3. Mathematical formulation .............................................................................................. 16 2.4. Solution Method............................................................................................................. 18 vii 2.4.1. Augmented Lagrangian Decomposition ................................................................ 19 2.5. Numerical Examples ...................................................................................................... 25 2.6. Summary ........................................................................................................................ 37 Chapter 3 ........................................................................................................................................ 40 Centralized – decentralized optimization for refinery scheduling ..................................... 40 3. 3.1. Introduction .................................................................................................................... 40 3.2. Problem Definition......................................................................................................... 42 3.3. Mathematical Formulation ............................................................................................. 44 3.4. Solution Approach ......................................................................................................... 51 3.5. Numerical Results .......................................................................................................... 56 3.6. Summary ........................................................................................................................ 62 Chapter 4 ........................................................................................................................................ 66 4. Scheduling of a large-scale oil-refinery operations ........................................................... 66 4.1. Introduction .................................................................................................................... 66 4.2. Problem Definition......................................................................................................... 71 4.2.1. 4.3. Mathematical Formulation ............................................................................................. 76 4.3.1. 4.4. Case study description ........................................................................................... 74 Valid Inequities ...................................................................................................... 88 Results on the Case Studies ........................................................................................... 90 4.4.1. Case study 1 ........................................................................................................... 92 4.4.2. Case Study 2 .......................................................................................................... 95 viii 4.5. Summary ........................................................................................................................ 99 Chapter 5 ...................................................................................................................................... 104 5. Lagrangian Decomposition Approach for Refinery Operations Scheduling ................... 104 5.1. Introduction .................................................................................................................. 104 5.2. Refinery Operations Scheduling .................................................................................. 108 5.3. Solution Strategy .......................................................................................................... 109 5.4. Lagrangian Relaxation Framework and Sub-problems................................................ 111 5.4.1. Relaxed Problem .................................................................................................. 112 5.4.1.1. Relaxed Blend Scheduling Problem ................................................................ 116 5.4.1.2. Relaxed Production Unit Scheduling Problem ................................................ 118 5.4.2. Construction of a Feasible Schedule .................................................................... 122 5.4.2.1. Restricted Relaxed Problem ............................................................................. 124 5.4.2.2. Upper Bounding Problem ................................................................................ 128 5.5. Strengthening the Lagrangian Relaxation .................................................................... 129 5.5.1. Preprocessing Step ............................................................................................... 129 5.5.2. Strengthening Lower Bound ................................................................................ 130 5.5.3. Construction of lower bound from the best upper bound .................................... 132 5.6. Lagrangian Decomposition Algorithm ........................................................................ 133 5.6.1. Updating the Multipliers ...................................................................................... 134 5.6.2. Stopping Criteria .................................................................................................. 135 5.6.3. Algorithm ............................................................................................................. 136 ix 5.7. Computational results .................................................................................................. 139 5.8. Summary ...................................................................................................................... 149 Chapter 6 ...................................................................................................................................... 156 6. Efficient Decomposition Approach for Large Scale Refinery Scheduling ...................... 156 6.1. Introduction .................................................................................................................. 156 6.2. Mathematical Formulations ......................................................................................... 157 6.3. Solution Strategy .......................................................................................................... 163 6.3.1. Blend Scheduling Problem................................................................................... 166 6.3.2. Production Unit Scheduling Problem .................................................................. 170 6.3.3. Restricted Blend Scheduling Problem ................................................................. 176 6.3.4. Integer Cuts .......................................................................................................... 180 6.4. Proposed Heuristic Algorithm ..................................................................................... 181 6.4.1. Termination Criterion .......................................................................................... 184 6.4.2. Evaluation of scheduling horizon ........................................................................ 184 6.5. Case Study ................................................................................................................... 185 6.6. Computational Results ................................................................................................. 187 6.7. Summary ...................................................................................................................... 201 Chapter 7 ...................................................................................................................................... 208 7. Augmented Lagrangian Relaxation Approach for Refinery Scheduling ......................... 208 7.1. Introduction .................................................................................................................. 208 7.2. Solution Strategy .......................................................................................................... 209 x 7.2.1. Augmented Lagrangian method ........................................................................... 209 7.2.2. Decomposition Strategy ....................................................................................... 211 7.3. Decomposed Models .................................................................................................... 216 7.3.1. Relaxed Blend Scheduling Problem .................................................................... 217 7.3.2. Relaxed Production Unit Scheduling Problem .................................................... 221 7.4. Strengthening the Augmented Lagrangian Decomposition ......................................... 226 7.4.1. Preprocessing Step ............................................................................................... 226 7.4.2. Strengthening Decomposed Problems ................................................................. 228 7.5. Augmented Lagrangian Optimization Algorithm ........................................................ 235 7.6. Examples ...................................................................................................................... 240 7.6.1. 7.7. Computational Results ......................................................................................... 243 Summary ...................................................................................................................... 256 Acknowledgement of Prior Works .............................................................................................. 263 Appendix...................................................................................................................................... 264 Appendix Chapter 2 ................................................................................................................. 264 Appendix Chapter 3 ................................................................................................................. 268 Appendix Chapter 6 ................................................................................................................. 277 Appendix Chapter 7 ................................................................................................................. 287 Bibliography ................................................................................................................................ 295 xi List of Tables Table 2.1 Model statistics of full-space integrated problem .......................................................... 26 Table 2.2 Computational results for example 1 ............................................................................. 27 Table 2.3 Computational results for example 2 ............................................................................. 33 Table 2.4 Computational results for example 3 ............................................................................. 35 Table 3.1 Characteristics of mathematical formulation using Scenario 1...................................... 57 Table 3.2 Characteristics of mathematical formulation using Scenario 2...................................... 58 Table 3.3 Scenario 1 Centralized vs. Decentralized ...................................................................... 59 Table 3.4 Scenario 2 Centralized vs. Decentralized ...................................................................... 60 Table 4.1 Price of raw materials and final products ...................................................................... 90 Table 4.2 Penalty parameters in objective function ....................................................................... 91 Table 4.3 Group A products demand order data for case study 1 .................................................. 93 Table 4.4 Group B products demands data for case study 1 .......................................................... 94 Table 4.5 Computational performance of case study 1 (without valid inequalities) ...................... 94 Table 4.6 Computational performance of case study 1 (with valid inequalities) ........................... 94 Table 4.7 Material flow in and out of multipurpose tank for case study 1, example 1 .................. 95 Table 4.8 Group A products demand order data for case study 2 .................................................. 96 Table 4.9 Group B products demands data for case study 2 .......................................................... 97 Table 4.10 Computational results for case study 2 (with valid inequalities) ................................. 97 Table 5.1 Model statistics for preprocessing problems................................................................ 139 Table 5.2 Model statistics ............................................................................................................ 140 Table 5.3 Computational Results ................................................................................................. 142 Table 5.4 Time spent (%) in each step compared to the total solution time of Lagrangian decomposition algorithms ............................................................................................................ 143 xii Table 6.1 Possible task assignment solutions based on current iteration task changeovers parameter i ,i, j ............................................................................................................................. 181 Table 6.2 Penalty parameters in objective function ..................................................................... 187 Table 6.3 Model statistics ............................................................................................................ 188 Table 6.4 Computational results .................................................................................................. 191 Table 6.5 Computational results .................................................................................................. 194 Table 6.6 Computational performance results ............................................................................. 195 Table 6.7 Algorithm progress and termination ............................................................................ 196 Table 7.1 Parameters in objective function .................................................................................. 243 Table 7.2 Model statistics for Example 1 .................................................................................... 245 Table 7.3 Numerical results for Example 1 ................................................................................. 247 Table 7.4 Model statistics for Example 2 .................................................................................... 249 Table 7.5 Numerical results for Example 2 ................................................................................. 250 Table 7.6 Model statistics for Example 3 .................................................................................... 252 Table 7.7 Numerical results for Example 3 ................................................................................. 253 Table 7.8 Model statistics for Example 5 .................................................................................... 254 Table 7.9 Numerical results for Example 5 ................................................................................. 255 Table A2-1 Example 1: Process data for production sites ........................................................... 264 Table A2-2 Production and Inventory cost data .......................................................................... 266 Table A2-3 Transportation and Backorder cost data ................................................................... 266 Table A3-1 Demand in Thousand barrels .................................................................................... 268 Table A3-2 Production rates in thousand barrels/hour ................................................................ 268 Table A3-3 Recipe data ............................................................................................................... 269 Table A3-4 Storage tank capacity data in thousand barrels ......................................................... 270 Table A6-1 Production rates in thousand barrels/hour ................................................................ 277 xiii Table A6-2 Minimum/maximum recipe data............................................................................... 278 Table A6-3 Storage tank capacity data in thousand barrels ......................................................... 280 Table A6-4 Group A products demand order data for the case study .......................................... 282 Table A6-5 Group B products demands data for the case study .................................................. 285 Table A7-1 Production rates (thousand barrels/hour) for a second diesel blend unit .................. 290 Table A7-2 Storage tank capacity data for blend component tanks............................................. 290 Table A7-3 Group B products demands data for the case study .................................................. 291 Table A7-4 Group A products demand order data for the case study .......................................... 292 xiv List of Illustrations Figure 1.1 Major elements of enterprise-wide optimization. ........................................................... 2 Figure 2.1 Multisite production and distribution network ............................................................. 11 Figure 2.2 Constraints matrix structure of an integrated multisite model ...................................... 19 Figure 2.3 Constraints matrix structure of a reformulated model .................................................. 20 Figure 2.4 Augmented Lagrangian-DQA decomposition algorithm ............................................. 25 Figure 2.5 Production facility state and task network (STN) representation (Example 1) ............ 26 Figure 2.6 Solution procedure of the ALO-DQA method (Example 1 and 90 planning periods) . 28 Figure 2.7 Production profiles of products (P1 and P2) at production sites (S1, S2, and S3) obtained using the ALO-DQA method (Example 1 and 30 planning periods) .............................. 29 Figure 2.8 Shipment profiles of products (P1 and P2) for production site 1 to markets (M1, M2, and M3) obtained using the ALO-DQA method (Example 1 and 30 planning periods) ............... 30 Figure 2.9 Shipment profiles of products (P1 and P2) for production site 2 to markets (M1, M2, and M3) obtained using the ALO-DQA method (Example 1 and 30 planning periods) ............... 30 Figure 2.10 Shipment profiles of products (P1 and P2) for production site 3 to markets (M1, M2, and M3) obtained using the ALO-DQA method (Example 1 and 30 planning periods) ............... 31 Figure 2.11 Sales profiles of products (P1 and P2) for markets (M1, M2, and M3) obtained using the ALO-DQA method (Example 1 and 30 planning periods) ...................................................... 31 Figure 2.12 Inventory profiles of products (P1 and P2) at production sites (S1, S2, and S3) obtained using the ALO-DQA method (Example 1 and 30 planning periods) .............................. 32 Figure 2.13 Production facility state and task network (STN) representation (Example 2) .......... 33 Figure 2.14 Solution procedure of the ALO-DQA method (Example 2, T = 90 periods) ............. 34 Figure 2.15 Solution procedure of the ALO-DQA method (Example 3, T = 90 periods) ............. 36 Figure 3.1 State and unit representation of Honeywell Hi-Spec problem ..................................... 43 Figure 3.2 Intermediate storage tank connecting two sub-systems................................................ 53 xv Figure 3.3 Decomposition of the refinery problem under consideration ....................................... 54 Figure 3.4 Gantt chart of the operation schedule for example 1, centralized system, scenario 1. . 61 Figure 3.5 Gantt chart of the operation schedule for Case 4, decentralized system, scenario 1. ... 62 Figure 4.1 Graphic overview of a standard refinery system .......................................................... 68 Figure 4.2 Schematic of the refinery production plant for reference example .............................. 75 Figure 5.1 Graphic overview of a standard refinery system: includes no blend components storage tanks. ............................................................................................................................................ 107 Figure 5.2 Spatial decomposition of a refinery operations network ............................................ 110 Figure 5.3 Framework of classical Lagrangian Decomposition Algorithm ................................. 112 Figure 5.4 Gantt chart for a blender receiving component s. ....................................................... 115 Figure 5.5. Algorithm for the proposed restricted-Lagrangian decomposition (restricted-LD) .. 138 Figure 5.6 Convergence of upper and lower bound of Lagrangian decomposition for example 1. ..................................................................................................................................................... 144 Figure 5.7 Convergence of upper and lower bound of Lagrangian decomposition for example 2. ..................................................................................................................................................... 145 Figure 5.8 Convergence of upper and lower bound of Lagrangian decomposition for example 3. ..................................................................................................................................................... 146 Figure 5.9 Convergence of upper and lower bound of Lagrangian decomposition for example 4. ..................................................................................................................................................... 147 Figure 5.10 Convergence of upper and lower bound of Lagrangian decomposition for example 5. ..................................................................................................................................................... 148 Figure 5.11 Convergence of upper and lower bound of Lagrangian decomposition for example 6. ..................................................................................................................................................... 149 Figure 6.1 Decomposition of an oil-refinery operations network by splitting blend components pipelines ....................................................................................................................................... 164 xvi Figure 6.2 Illustration of scheduling sub-problems obtained by splitting component flow streams connecting process units directly to blend headers ...................................................................... 164 Figure 6.3 Proposed solution approach. ....................................................................................... 166 Figure 6.4 Determining parameters for PSP using a solution of a blend header j ....................... 169 Figure 6.5 Determining parameter tupo for PSP using the corresponding blend header j ........... 169 Figure 6.6 Proposed heuristic algorithm ...................................................................................... 184 Figure 6.7 Schedule evaluation of BSP, PSP, and RBSP. Here i represents index for blend task. ..................................................................................................................................................... 185 Figure 6.8 Gantt chart for Example 5 obtained using integrated full-scale model ...................... 199 Figure 6.9 Gantt chart for Example 5 obtained from the heuristic algorithm .............................. 200 Figure 7.1 Spatial decomposition of a refinery network .............................................................. 212 Figure 7.2 Spatial decomposition of blend sub-network structure (BN) ..................................... 217 Figure 7.3 Production unit operations network ............................................................................ 222 Figure 7.4. Augmented Lagrangian optimization algorithm ........................................................ 240 Figure 7.5 Solution procedure for Example 1, problem 5. .......................................................... 248 Figure 7.6 Solution procedure for Example 2, problem 5 ........................................................... 251 Figure 7.7 Solution procedure for Example 3, problem 5. .......................................................... 254 Figure 7.8 Solution procedure for Example 4, problem 5. .......................................................... 256 Figure A2-1 Demand data for Example 1 .................................................................................... 267 Figure A3-1 Gantt chart of tank loading schedule for example 1, Centralized system, scenario 1 ..................................................................................................................................................... 271 Figure A3-2 Gantt chart of the tank unloading schedule for example 1, Centralized system, scenario 1 ..................................................................................................................................... 272 Figure A3-3 Gantt chart of tank loading schedule for example 1, Decentralized system, scenario 1 ..................................................................................................................................................... 272 xvii Figure A3-4 Gantt chart of the tank unloading schedule for example 1, Decentralized system, scenario 1 ..................................................................................................................................... 273 Figure A3-5 Gantt chart of the operation schedule for example 1, Centralized system, scenario 2 ..................................................................................................................................................... 273 Figure A3-6 Gantt chart of the operation schedule for example 1, Decentralized system, scenario 2 ................................................................................................................................................... 274 Figure A3-7 Gantt chart of tank loading schedule for example 1, Centralized system, scenario 2 ..................................................................................................................................................... 274 Figure A3-8 Gantt chart of tank unloading schedule for example 1, Centralized system, scenario 2 ..................................................................................................................................................... 275 Figure A3-9 Gantt chart of tank loading schedule for example 1, Centralized system, scenario 2 ..................................................................................................................................................... 275 Figure A3-10 Gantt chart of tank unloading schedule for example 1, Centralized system, scenario 2 ................................................................................................................................................... 276 xviii 1 Chapter 1 1. Introduction 1.1. Enterprise Wide Optimization The process industry is a key sector in the global economy, converting raw materials such as crude oil, water, and natural resources into thousands of products. According to the American Chemistry Council, over 96% of all manufactured goods are dependent on chemical industry and U.S. produces over 15% of world‟s chemical output, amounting to US$812 billion in (2014). In particular, the petroleum refining industry is the largest source of energy products in the world and is supplying about 39% of total U.S. energy demand and 97% of transportation fuels. Process industry has grown increasingly complex in the last 20 years as a result of tighter competition, stricter environmental regulations, uncertainty in the prices of energy, raw materials, and products, and lower-margin profits. Globalization trends of past few decades have significantly increased the scale and complexity of the modern enterprise by transforming them into global network consisting of multiple business units and functions. Today process industries involve multipurpose, multisite production facilities producing hundreds of products, located in different regions and countries and servicing international market. (Wassick, 2009) In last decades, enterprises are realizing the importance of enterprise wide optimization in reducing the overall costs and remaining competitive in a dynamic global marketplace. Enterprise-wide optimization involves coordinated optimization of the operations of supply, manufacturing, and distribution; and integration of the information and decisions-making among the various functions that comprise the supply chain of the company. (I. Grossmann, 2005; I. E. Grossmann, 2014) The concept of enterprise-wide optimization (EWO) lies at the interface of the process system engineering, operations research and supply chain optimization and these concepts are suitably positioned to provide decisions making models and algorithms for 2 optimization of an integrated manufacturing and distribution complexes. Supply chain optimization can be considered an equivalent term for describing the enterprise-wide optimization although supply chain optimization places more emphasis on logistics and distribution, whereas enterprise-wide optimization is aimed at manufacturing facilities optimization. (Shapiro, 2001) The process industry supply chains vary from the petroleum supply chain (N. K. Shah et al., 2011) to pharmaceutical industry supply chain (Nilay Shah, 2004); however they all include manufacturing as a major component according to I. E. Grossmann (2014); Wassick (2009). The main goal of EWO is to maximize profits while minimizing costs, inventories and to this ends, major operational activities of EWO is planning, scheduling and real-time optimization and control (Figure 1.1). Figure 1.1 Major elements of enterprise-wide optimization. In the field of enterprise-wide optimization, planning and scheduling are the most important operational decisions. Objectives of the planning and scheduling problems is to determine the allocation of available resources over time to perform a set of tasks required to manufacture one or more products as to satisfy global demand. The long to medium term planning covers a time horizon of between few months to a year and is concerned with decisions such as production, inventory, and distribution profiles. Short-term scheduling decision deals with issues such as assignment of tasks to units and sequencing of tasks in each unit and typically covers time horizon of between days to a few weeks. Typically, much of the decision-making in a 3 supply chain is focused across solving sub-problems as an entity, but from the enterprise-wide performance viewpoint, local improvements at any sublevel do not necessary lead to an overall improvement and to realize full potential of EWO, integrated approach is necessary. Amongst the challenge involve in EWO in process industry, the chief challenge is coordinated decisionmaking across different functions in industry (procurement, manufacturing, distributions), between various geographically distributed manufacturing sites, and across three levels of decisions-making process. (Shapiro, 2001) While the first challenge relates to spatial integration and the third challenge involves temporal integration across different timescales. Coordinated decision making between geographically distributed integrated manufacturing sites deals with not only spatial integration but also with temporal integrations. Several of the aforementioned issues can be addressed in part through integration of planning and scheduling decision-making for multi-product, multi-site production and distribution facilities. 1.1.1. Problems and Challenges I. Grossmann (2005) discussed four major challenges in application of EWO: (a) Mathematical modeling, (b) multiscale optimization, (c) optimization under uncertainty, and (d) algorithmic and computational challenge. The first challenge involves development of production and scheduling models that can effectively capture the complexity of various operations but can also be solved in reasonable time frame. Second challenge involves difficulty associated with coordination between different time scale and over different geographically located sites. Uncertainty is inherent in supply chain (e.g. prices, demand, equipment breakdown) and effectively addressing it can effect industry profits. However, before addressing uncertainties, computationally effective deterministic models should be developed. Various models developed in previous three points, are large scale complex MILP or MINLP models and they require the application of various decomposition techniques. Comprehensive reviews of challenges faced in EWO of process industry are presented by I. E. Grossmann (2014); Nilay Shah (2004, 2005). 4 Scheduling models in process industry address continuous, batch, and semi-continuous production systems and complexity involved modeling these three different production system vary. In continuous production units, there is a simultaneous inlet and outlet streams, where as in batch process, simultaneous inlet and outlet streams are not allowed. Large varieties of products are produced using batch processes and food, beverages, pharmaceutical products, paint, fertilizer, and cement are a few of the categories of products produced using batch processes. Oil refinery is one of the prominent systems with continuous production process. In last ten years, great progress has been in short-term batch scheduling process, few of these works are Burkard et al. (2002); Pedro M. Castro et al. (2009); P. M. Castro et al. (2011); He and Hui (2007); Ierapetritou and Floudas (1998 ); Janak et al. (2006); Janak et al. (2004); Kondili et al. (1993); Maravelias and Grossmann (2003); Moniz et al. (2014). The challenge in batch process industry lies in developing effective solution methodology for integrating existing scheduling models with planning level model and integrating decisions making across multisite production and distribution facilities. Whereas, the main challenge in continuous process industry lies in developing general purpose scheduling models. Every continuous process industry has its own unique problems and arriving at one scheduling model for different types of continuous process operations is difficult and thus focus is on development of general scheduling models for specific industry. (I. Grossmann, 2005) In petroleum industry, different operations in refinery have their own scheduling model instead of an integrated scheduling model for overall refinery operations. Before tackling planning and scheduling integration in refinery, efficient and effective scheduling model for overall efficient refinery operations need to be developed. Furthermore, refinery process is very complex and their scheduling models are large scale complex mixed integer models that require novel decomposition strategies.(Kelly & Zyngier, 2008; Shaik & Floudas, 2007) 5 1.2. Planning and Scheduling in Process Industry In the recent years, the area of integrated planning and scheduling has received much attention for single-site batch production facilities. However, the current manufacturing environment for process industry has changed from a traditional single-site production plant to a more integrated global production serving the emerging market. (Wassick, 2009) Modern process industries operate as a large integrated complex that involve multi-product, multi-purpose, and multi-site facilities serving a global market. The process industries supply chain can be defined to be composed of production facilities and distribution centers, where final products produced at production facilities are transported to distribution center to satisfy the customers demand. In current global market, spatially distributed production facilities across various geographical locations can no longer be regarded as isolated from each other and interactions between the production facilities and the distribution centers should be taken into account when making decisions. In this context, the issues of enterprise planning and coordination across production plants and distribution facilities are important for robust response to global demand and to maintain business competiveness, sustainability, and growth. (Papageorgiou, 2009) As the pressure to reduce the costs and inventories increases, centralized approaches have become the main policies to address supply chain optimization. (Grossmann, 2005) The integrated planning and scheduling model for multi-site facilities is important to ensure the consistency between planning and scheduling level decisions and to optimize production and transportation costs. Wassick (2009) proposed a planning and scheduling model based on resource task network for an integrated chemical complex. He considered the enterprise-wide optimization of the liquid waste treatment network with their model. Kreipl and Pinedo (2004) discussed issues present in modeling the planning and scheduling decisions for supply chain management. For a multisite facilities, the size and level of interdependences between these sites present unique challenges to the integrated tactical production planning and day-to-day scheduling problem and these 6 challenges are highlighted by Kallrath (2002b). For further elucidation of various aspects of planning, reader is directed to the work of Timpe and Kallrath (2000) and Kallrath (2002a). 1.3. Refinery Operations Scheduling The continuous process plants scheduling have drawn less consideration in the literature compared to that of batch plants even though continuous plants are prevalent in the chemical process industries. Of the continuous process industries, the oil refinery production operation is one of the most complex chemical industries, which involves many different and complicated processes with various connections. Instead of tackling a comprehensive large-scale refinery operations optimization problem, decomposition approaches are generally exploited. Oil refinery manufacturing operations can be decomposed into three problems: (1) crude-oil unloading, mixing, and inventory control, (2) scheduling of production units, and (3) finish products blending and distribution. (Jia & Ierapetritou, 2004) The goal of EWO is to optimized overall refinery operations in a coordinated fashioned. Depending on the problem characteristics as well as the required flexibility in the solution, scheduling models can be based on either a discrete, a continuous, or hybrid time domain representation. (Iiro Harjunkoski et al., 2014; J. Li & Karimi, 2011; Mouret et al., 2010; Neiro et al., 2014) Real-life features such as multipurpose production units, multipurpose product tanks, parallel non-identical blenders, minimum run lengths, sequence dependent changeovers, product giveaway, piecewise constant profiles for blend component qualities and feed rates, etc. introduces in more operational constraints and many combinatorial decisions, that renders this large scale mixed integer problem difficult to solve without decomposition solution strategies. (Kelly, 2006; Kelly & Zyngier, 2008; Shaik & Floudas, 2007; Shaik et al., 2009) 7 1.4. Motivation: development and implementation of decomposition tools for planning and scheduling in process industry Despite the many of the potential benefits in coordinated decisions making in EWO, aforementioned challenges in coordinated decisions making in planning and scheduling for multiproduct, multisite product and distribution facilities and scheduling of overall refinery operations, concepts of enterprise-wide optimization are underutilized. This is due to the difficulties associated with building effective mathematical formulation that captures real world complexity without becoming incomprehensive to solve and for efficient models to be utilized for real world application, the challenge lies in developing decomposition approaches for large scale mixed integer models.(I. E. Grossmann, 2014) First objective of this dissertation is to develop mathematical formulation for multisite batch-process production and distribution facility planning and propose decomposition approach to address multi-scale optimization problem arising from integration of planning and scheduling decisions level. Second objective of this dissertation is to propose scheduling models for large scale refinery operations, to develop efficient decomposition based methodology to address large scale optimization problems and demonstrate applicability of decomposition methodologies. The methodologies to tackle large scale optimization models include, valid inequalities, heuristic algorithm, Lagrangian relaxation decomposition, and augmented Lagrangian optimization. Each of these decomposition methods will be discussed in the context of refinery operations scheduling and demonstrated using case studies related to real refinery complex. 1.5. Outline of dissertation Each of the six main chapters in this dissertation will emphasize a specific concept or tool for enterprise-wide optimization. Chapter 2 will present decomposition based methodology to solve full space integrated planning and scheduling problem for multisite, multiproduct batch production plants and multisite distribution facilities. Chapter 3 will present spatial 8 decomposition strategy for refinery operations where centralized and decentralized decision making process is compared and conditions during which decentralized approach gives global optimal solution are discussed. A unified comprehensive refinery operation scheduling model that incorporates many of the real world operational and logistics rules is presented in Chapter 4. The resulting model is large scale mixed integer linear programming model (MILP) and valid inequalities are developed to reduce complexity of the model by reducing the total number of nodes and iterations in branch and bound framework. Refinery operations scheduling model incorporating many logistics rules is difficult to solve to optimality in reasonable computational time even after inclusion of valid inequalities. For refinery without any blend component tanks, Chapter 5 tackles complexity of large scale scheduling model by developing mathematical decomposition, Lagrangian relaxation, while Chapter 6 focuses on efficient heuristic decomposition algorithm. In Chapter 7, a general augmented Lagrangian decomposition for different refinery operations configuration (rundown streams, blend component tanks, or both) is introduced and applied to number of case studies to illustrate its effectiveness. 9 Chapter 2 2. Planning and scheduling of multisite batch production and distribution facilities The current manufacturing environment for process industry has changed from a traditional single-site, single market to a more integrated global production mode where multiple sites are serving a global market. In this chapter, the integrated planning and scheduling problem for the multisite, multiproduct batch plants is considered. The major challenge for addressing this problem is that the corresponding optimization problem becomes computationally intractable as the number of production sites, markets, and products increases in the supply chain network. To effectively deal with the increasing complexity, the block angular structure of the constraints matrix is exploited by relaxing the inventory constraints between adjoining time periods using the augmented Lagrangian decomposition method. To resolve the issues of non-separable crossproduct terms in the augmented Lagrangian function, diagonal approximation method is applied. Several examples have been studied to demonstrate that the proposed approach yields significant computational savings compared to the full-scale integrated model. 2.1. Introduction Modern process industries operate as a large integrated complex that involve multiproduct, multipurpose, and multisite production facilities serving a global market. The process industries supply chain is composed of production facilities and distribution centers, where the final products are transported from the production facilities to distribution centers and then to retailers to satisfy the customers demand. In current global market, spatially distributed production facilities across various geographical locations can no longer be regarded as independent from each other and interactions between the manufacturing sites and the distribution centers should be taken into account when making decisions. In this context, the issues of enterprise planning and coordination across production plants and distribution facilities are important for robust response to global demand and to maintain business competiveness, sustainability, and growth. 10 (Papageorgiou, 2009) As the pressure to reduce the costs and inventories increases, centralized approaches have become the main policies to address supply chain optimization. An excellent overview of the enterprise-wide optimization (EWO) and the challenges related to process industry supply chain is highlighted by I. Grossmann (2005). Varma et al. (2007) described the main concepts of EWO and presented the potential research opportunities in addressing the problem of EWO models and solution approaches. Supply chain optimization can be considered an equivalent term for describing the enterprise-wide optimization according to Shapiro (2001) although supply chain optimization places more emphasis on logistics and distribution, whereas enterprise-wide optimization is aimed at manufacturing facilities optimization. Key issues and challenges faced by process industry supply chain are highlighted by Nilay Shah (2004, 2005). Traditional supply chain management planning decisions can be divided into three levels: strategic (long-term), tactical (medium-term), and operational (short-term). The long-term planning determines the infrastructure (e.g. facility location, transportation network). The medium-term planning covers a time horizon between few months to a year and is concerned with decisions such as production, inventory, and distribution profiles. Finally, short-term planning decision deals with issues such as assignment of tasks to units and sequencing of tasks in each unit. The short-term planning level covers time horizon between days to a few weeks and at production level, is typically refer to as scheduling. Wassick (2009) proposed a planning and scheduling model based on resource task network for an integrated chemical complex. He considered the enterprise-wide optimization of the liquid waste treatment network with their model. Kreipl and Pinedo (2004) discussed issues present in modeling the planning and scheduling decisions for supply chain management. For a multisite facilities, the size and level of interdependences between these sites present unique challenges to the integrated tactical production planning and day-to-day scheduling problem and 11 these challenges are highlighted by Kallrath (2002b). For further elucidation of various aspects of planning, the reader is directed to the work of Timpe and Kallrath (2000) and Kallrath (2002a). A simple network featuring the multisite facilities is given in Figure 2.1, where multiple products may be produced in individual process plants at different locations spread across geographic region and then transported to distribution centers to satisfy customers demand. These multisite plants produce a number of products driven by market demand under operating conditions such as sequence dependent switchovers and resource constraints. Each plant within the enterprise may have different production capacity and costs, different product recipes, and different transportation costs to the markets according to the location of the plants. To maintain economic competitiveness in a global market, interdependences between the different plants, including intermediate products and shared resources need be taken into consideration when making planning decisions. Furthermore, the planning model should take into account not only individual production facilities constraints but also transportation constraints because in addition to minimizing the production cost, it‟s important to minimize the costs of products transportation from production facilities to the distribution center. Thus, simultaneous planning of all activities from production to distribution stage is important in a multisite process industry supply chain. (N. Shah, 1998) Figure 2.1 Multisite production and distribution network 12 Wilkinson et al. (1996) proposed an aggregated planning model based on the resource task network framework developed by Pantelides (1994). Their proposed planning model considers integration of production, inventory, and distribution in multisite facilities. Lin and Chen (2007) developed a multistage, multisite planning model that deals with routings of manufactured products demand among different production plants. They simultaneously combine two different time scales (i.e. monthly and daily) in their formulation by considering varying time buckets. Verderame and Floudas (2009) developed an operational planning model which captures the interactions between production facilities and distribution centers in multisite production facilities network. Their proposed multisite planning with product aggregation model (Multisite-PPDM) incorporates a tight upper bound on the production capacity and transportation cost between production facilities and customers distribution centers in the supply chain network under consideration. A multisite production planning and distribution model is proposed by Jackson and Grossmann (2003) where they utilized nonlinear process models to represent production facilities. They have exploited two different decomposition schemes to solve the large-scale nonlinear model using Lagrangian decomposition. In temporal decomposition, the inventory constraints between adjoining time periods are dualized in order to optimize the entire network for each planning time period. In spatial decomposition technique, interconnection constraints between the sites and markets are dualized in order to optimize each facility individually. They conclude that temporal decomposition technique performs far better than spatial decomposition technique. The traditional strategy to address planning and scheduling level decisions is to follow a hierarchical approach in which planning decisions are made first and then scheduling decisions are made using planning demand targets. However, this approach does not consider any interactions between the two decision making levels and thus the planning decisions may result in suboptimal or even infeasible scheduling problems. Due to significant interactions between planning and scheduling decisions levels in order to determine the global optimal solution it is 13 necessary to consider the simultaneous optimization of the planning and scheduling decisions. However, this simultaneous optimization problem leads to a large problem size and the model becomes intractable when typical planning horizon is considered. For an overview of issues, challenges and optimization opportunities present in production planning and scheduling problem, the reader is referred to the work of Maravelias and Sung (2009). In recent years, the area of integrated planning and scheduling for single site has received much attention. Different decomposition strategies are developed to effectively deal with a large scale integrated model. One of the existing approaches follows a hierarchical decomposition method, where the upper level planning problem provides a set of decisions such as production and inventory targets to the lower level problem to determine the detailed schedule. If the solution of lower level problem is infeasible, an iterative framework is used to obtain a feasible solution. (Bassett et al., 1996) To further improve this approach, tight upper bounds on production capacity are implemented in upper level problem in presence of an approximate scheduling model or aggregated capacity constraints. (Nilay Shah, 2005; Shapiro, 2001) Another related idea is the one that follows a hierarchical decomposition within a rolling horizon framework. In this model detailed scheduling models are used for a few early periods and aggregated models are used for later periods. (Dimitriadis et al., 1997; Z. Li & Ierapetritou, 2010b; Verderame & Floudas, 2008; Wu & Ierapetritou, 2007) A different decomposition strategy is based on the special structure of the large-scale mathematical programming model. The integrated planning and scheduling model has a block angular structure which arises when a single scheduling problem is used over multiple planning periods. The constraints matrix of the integrated problem has complicating variables that appear in multiple constraints. By making copies of the complicating variables, the complicating variables are transformed into complicating constraints (linking constraints) and these complicating constraints can be relaxed using the Lagrangian relaxation method. One major drawback of the Lagrangian relaxation (LR) is that there is duality gap between the solution of 14 the Lagrangian relaxation method and original problem and to resolve this issue, augmented Lagrangian relaxation (ALR) method should be used Y. Li et al. (2008); Tosserams et al. (2006); Tosserams et al. (2008). Z. Li and Ierapetritou (2010a) applied augmented Lagrangian optimization method to integrated planning and scheduling problem for single site plants. One disadvantage of ALR method is the non-separability of the relaxed problem which arises due to the quadratic penalty terms present in the objective function. To resolve the issue of the nonseparability, Z. Li and Ierapetritou (2010a) studied two different approaches. The first approach is based on linearization the cross-product terms using diagonal quadratic approximation (DQA). (Y. Li et al., 2008) However, in this approach, an approximation of the relaxation problem is solved and it may not lead to a global optimal solution of the original problem. In the second approach, Z. Li and Ierapetritou (2010a) proposed a two-level optimization method which solves an exact relaxation problem. However, the proposed two-level optimization strategy requires a non-smooth quadratic problem to be optimized at every iteration. They conclude that DQA-ALO method is more effective than the two-level optimization method for the integrated planning and scheduling problems. Even though most companies operate in a multisite production manner, very limited attention has been paid on integrating planning and scheduling decisions for multisite facilities. The integrated planning and scheduling model for multisite facilities is important to ensure the consistency between planning and scheduling level decisions and to optimize production and transportation costs. Since the production planning and scheduling level deals with different time scales, the major challenge for the integration using mathematical programming methods lies in addressing large scale optimization models. The full-scale integrated planning and scheduling optimization model spans the entire planning horizon of interest and includes decisions regarding all the production sites and distribution centers. When typical planning horizon is considered, the 15 integrated full-scale problem becomes intractable and a mathematical decomposition solution approach is necessary. In this work, augmented Lagrangian relaxation method is applied to solve the multisite production and distribution optimization problem. The chapter is organized as follow. The problem statement is given in Section 2.2, whereas Section 2.3 presents the problem formulation. The general augmented Lagrangian method and its application to multisite facility is given in Section 2.4. The results of examples studied are shown in section 2.5 and the chapter concludes with summary in section 2.6. 2.2. Problem Statement The supply chain network (Figure 2.1) under investigation contains multiple batch production facilities which supply products to multiple distribution centers. Every production site may supply all distribution centers but all the products may not be produced at every production site. In the proposed model, it is assumed that one cannot sell more products to a market than the market forecasted demand. Thus the requested demand acts as an upper bound on finished product sales. The proposed model tries to satisfy the market demands; however it allows for unsatisfied demand to be carried over to the next planning period (backorder) and also allows for partial order fulfillment. The unsatisfied demand and backorders are penalized on a daily basis in order to maximize the degree of customer demand order fulfillment. Following assumptions are made: unlimited supply of raw materials is available and fixed and variable production, storage, and backorder costs are known for the planning horizon under consideration. The transportation costs are also assumed to be known. It is further assumed that there are no shipping delays in the network and the length of time of the planning horizon is such that the effects of transportation delays are neglected. Given the daily demand profiles for each distribution center, the goals of the integrated planning and scheduling problem is to ascertain the daily production target profiles for each 16 production facilities and product shipment profiles from production facilities to distribution centers so that demand is satisfied over the planning horizon under considerations (several months up to a year). The objective of the integrated problem is to minimize inventory, backorder, transportation, and production costs. 2.3. Mathematical formulation The multisite model includes production site constraints and distribution center (market) constraints. The set of products s ∈ PR are to be produced at various production sites (p ∈ PS) and are to be distributed to a global market (m ∈ M) over planning horizon (t ∈ T). To formulate the integrated planning and scheduling model, an integrated modeling approach is proposed in which planning and scheduling decisions level constraints are incorporated. The planning horizon is discretized into fixed time length (daily production periods) and for each planning period, a detailed scheduling model for each batch production facilities is considered. The detailed scheduling model is based on continuous time representation and notion of event points. (Ierapetritou & Floudas, 1998 ) The planning and scheduling decisions levels are inter-connected via production and inventory constraints. The integrated planning and scheduling model for a single-site proposed by Z. Li and Ierapetritou (2010a) is extended to accommodate multisite production facilities serving multiple markets. The extended integrated multisite model is as follows. min h Inv u U p s t p m s p ,t s p sS f t m m sS f m ,t s t m p d sp ,m Dsp ,m,t sS fp S mf FixCostip wvip, ,jt,n VarCostip bip, j,t,n t s.t. (1a) p jJ p iI jp n Invsp ,t Invsp ,t -1 Ps p ,t - D p , m ,t s , s S fp , p PS , t T (1b) , s S mf , m M , t T (1c) m U sm ,t U sm ,t -1 Demsm ,t - D pPs p , m ,t s 17 stsp,n,t N - stinsp ,t Ps p ,t , s S fp , p PS , t T (1d) stinsp ,t Invsp ,t -1 , s S fp , p PS , t T (1e) wv (1f) p ,t i , j ,n 1, j J p , n N , p PS , t T iI jp p ,t p ,t max p ,t p p vimin , j , p wvi , j , n bi , j , n vi , j , p wvi , j , n , i I j , j J , n N , p PS , t T (1g) stsp,n,t stcapsp , s S p , n N , p PS , t T (1h) stsp,n,t stsp,n,t1 b c, p s ,i iI sp stsp,n,t1 stinsp ,t p ,t i , j ,n b c, p s ,i iI sp b p, p s ,i iI sp jJ ip p ,t i , j , n 1 , p ,t i , j ,n 1 , s S p , n N , p PS , t T jJ ip s S p , p PS , t T jJ ip (1i) (1j) Tfi ,pj,,tn Tsip, ,jt,n ip, j wvip, j,t,n i p, j bip, j,t,n , i I jp , j J p , n N , p PS , t T (1k) Tsip, ,jt,n 1 Tfi ,pj,,tn H 1 wvip, j,t,n , i I jp , j J p , n N , p PS , t T (1l) (1m) Tsip, ,jt,n 1 Tf ip, ,jt,n H 1 wvip,,jt,n , i, i I jp , i i, j , j J p , n N , p PS , t T (1n) Tsip, ,jt,n 1 Tsip, ,jt,n , i I jp , j J p , n N , p PS , t T (1o) Tfi ,pj,,tn 1 Tfi ,pj,,tn , i I jp , j J p , n N , p PS , t T (1p) Tsip, ,jt,n H , i I jp , j J p , n N , p PS , t T (1q) Tfi ,pj,,tn H , i I jp , j J p , n N , p PS , t T (1r) Tsip, ,jt,n 1 Tfip, ,jt,n H 1 wvip,,jt ,n , i I jp , i I jp , j J p , n N , p PS , t T 18 The objective function shown in equation (1a) minimizes the total costs of the integrated model, which includes variable inventory costs, backorder costs, transportation costs, and production costs and fixed production costs. The planning level is modeled by constraints (1b1c). Equation 1b predicts the production targets ( Ps p ,t ), inventory targets ( Invsp ,t ), and shipping targets ( Dsp ,m ,t ) for each product. The constraint describing each distribution center (market) is given in equation (1c). As shown in equation (1c), the backorder balance is performed for each customer market by considering the demand forecast ( Demsm ,t ) of that market and the sales all the shipments of the product from one or combination of all production sites to that market ( Dsp ,m ,t ). Constraints (1d) assign production targets ( Ps p ,t ) of planning level solutions to scheduling level problem for each product to each production facility for different planning periods. Equation (1e) represents the connection between the inventory level requirements for the scheduling problems to that of the different planning periods for each product. In addition to constraints (1a-1e), the model also includes detailed scheduling constraints (1f-1r) for each production site (p ∈ PS) and for each planning period (t ∈ T). These scheduling level constraints are allocation constraints (1f), production capacity constraints (1g), storage capacity constraints (1h), material balance constraints (1i) and (1j), and sequence constraints (1k-1r). Equations (1a-1r) comprise the complete multisite batch production facilities and multisite distribution centers considering planning and scheduling decisions. 2.4. Solution Method The full-scale integrated model gives rise to a large scale optimization problem which requires the use of decomposition methods to be solved effectively. The appropriate mathematical decomposition approach is decided by analyzing the constraints matrix of the full-scale model. If the planning decision variables ( Invsp ,t , Ps p ,t , Dsp ,m ,t ,U sm ,t ) are denoted as X t , p and scheduling decision variables as Y t , p , then the structure of the integrated model can be illustrated as shown in 19 a constraints matrix (Figure 2.2). As it is seen in the Figure 2.2, the matrix has a block angular structure and these blocks are linked through the planning decisions variables, inventory and production targets for each production facilities ( Invsp ,t , Ps p ,t ). These complicating variables can be handled using augmented Lagrangian relaxation method described in the next section to obtain a decomposable structure. Figure 2.2 Constraints matrix structure of an integrated multisite model 2.4.1. Augmented Lagrangian Decomposition In order to obtain a decomposable structure, the complicating variables need to be transformed into complicating constraints and then, the model can be relaxed by eliminating complicating constraints from the total constraints set. The first step in obtaining the relaxation problem is to duplicate the planning inventory and production targets variables, using different variables in planning and scheduling problems and incorporate the coupling constraints (2f-2g) into the fullscale model. The production and inventory scheduling targets constraints are rewritten as (2q-2r). 20 This transforms the constraints matrix (Figure 2.2) into the matrix with complicating constraints shown in Figure 2.3. Figure 2.3 Constraints matrix structure of a reformulated model min h Inv u U p s t t m sS mf FixCost t p p ,t i wvi , j , n t s.t. m m ,t s s p ,t s p sS fp m p VarCostip bip, j,t,n p iI p jJ ip n Invsp ,t Invsp ,t -1 Ps p ,t - D p , m ,t , s d sp ,m Dsp ,m,t sS fp S mf (2a) s S fp , p PS , t T (2b) s S mf , m M , t T (2c) m U sm ,t U sm ,t -1 Demsm ,t - D p , m ,t , s pPs stsp,n,t N - stinsp ,t PPsp ,t , s S fp , p PS , t T (2q) stinsp ,t II sp ,t -1 , s S fp , p PS , t T (2r) II sp ,t Invsp ,t , s S fp , p PS , t T (2f) 21 PPsp ,t Psp ,t , s S fp , p PS , t T (2g) y p,t Y , p PS , t T (2h) The complicating constraints (2f-2g) link the planning level constraints with the scheduling level constraints. The constraints (2h) express a compact representation of the scheduling level constraints (1f-1r) for each production site and planning period. The reformulated model (2) is still not decomposable since the planning and scheduling problems are interconnected via coupling constraints thus the Augmented Lagrangian relaxation method is applied by dualizing the complicating constraints in equations (2f-2g), which involves removing them from the reformulated model constraints set and adding them to the objective function multiplied by the Lagrange multipliers ( sp ,t , sp ,t ) and quadratic penalty parameters (σ) as shown in equation (3a). Constraints (3a-3h) correspond to the augmented Lagrangian relaxation problem. f , , = min h Inv u U p s t t m p p ,t i wvi , j , n t p iI p p j J i n Inv p ,t s p ,t s p sS fp Invsp ,t Invsp ,t -1 Ps p ,t - - II sp ,t t D p , m ,t , s m ,t s m sS f FixCost t s.t. p m s p ,t s p sS f m t d sp ,m Dsp ,m,t p sS fp S mf VarCostip bip, j,t,n sp ,t Psp ,t - PPsp ,t P t p ,t s p sS fp - PPsp ,t p sS f Inv 2 p ,t s - II sp ,t (3a) 2 s S fp , p PS , t T (3b) s S mf , m M , t T (3c) m U sm ,t U sm ,t -1 Demsm ,t - D p , m ,t , s pPs stsp,n,t N - stinsp ,t PPsp ,t , s S fp , p PS , t T (3q) stinsp ,t II sp ,t -1 , s S fp , p PS , t T (3r) 22 y p,t Y , p PS , t T (3h) To improve the convergence to the feasible solution and to avoid duality gap that may result with just the Lagrangian terms ( sp ,t Ps p ,t - PPs p ,t and sp ,t Invsp ,t - II sp ,t ), the quadratic penalty term ( Inv 2 Psp ,t - PPsp ,t p ,t s - II sp ,t ) 2 is applied to the relaxation formulation as given in (3a). However, the quadratic penalty term in the objective function of the relaxation problem has nonseparable bilinear terms Psp,t PPsp,t and Invsp ,t II sp ,t . To resolve the non-separabilility issue, the diagonal quadratic approximation (DQA) method is applied to linearize the cross-product terms p ,t p ,t p ,t Inv p ,t s II sp ,t p ,t around the tentative solution P s , PP s , Inv s , II s as shown in the following equations ((Y. Li et al., 2008)). Inv II sp ,t PPsp ,t p ,t s P s p ,t p ,t 2 2 Invsp ,t II s 2 Psp,t PP s P p ,t 2 p ,t s PPsp,t Inv 2 P 2 p ,t s p ,t s p ,t 2 II s p ,t 2 PP s The objective function (3a) can be thus be rewritten in decomposable form given by equation (4a‟). f ( , , ) = f pl f p ,t sc (4a‟) tT pPS where, the f pl represents the objective function of the planning problem (4a, 4b, 4c). f pl , , min P , Inv , D ,U h Inv u U p s t p sS fp sp ,t Ps p ,t p sS f t t m s p ,t s t t m ,t s m sS mf t m d sp ,m Dsp ,m,t p sS fp S mf p ,t p ,t s Invs (4a) p sS fp p ,t p p Psp,t - PP s sS f 2 p ,t 2 Invsp ,t - II s p ,t p ,t 2 P s - PP s p ,t p ,t 2 Inv s - II s 23 s.t. D Invsp ,t Invsp ,t -1 Ps p ,t - p , m ,t , s s S fp , p PS , t T (4b) s S mf , m M , t T (4c) m U sm ,t U sm ,t -1 Demsm ,t - D p , m ,t , s pPs where f scp ,t represents the objective function of the scheduling sub-problems (5a, 5b-5d). The scheduling sub-problem is defined at each production site and for each planning period. f scp ,t , , sS fp s.t. min PP p ,t , II p ,t , y p ,t p ,t p ,t s II s FixCostip wvip, ,jt,n VarCostipbip, j,t,n sp,t PPsp,t iI p p j J i p ,t P s - PPsp ,t sS fp sS fp n 2 p ,t Inv s - II sp ,t 2 (5a) stsp,n,t N - stinsp ,t PPsp ,t , s S fp , p PS , t T (5b) stinsp ,t II sp ,t -1 , s S fp , p PS , t T (5c) y p,t Y , p PS , t T (5d) These quadratic problems are solved using a general augmented Lagrangian optimization and diagonal quadratic approximation (ALO-DQA) algorithm which is outlined in Figure 2.4. The ALO-DQA algorithm can provide an optimal solution if the objective functions and constraints function are convex, feasible region is bounded and closed, and the step size is sufficiently small. The algorithm has the following parameters: the initial Lagrange multipliers ( sp ,t ,0 , sp ,t ,0 ) which are chosen to be zero, the initial penalty parameter ( 0 1 ), and the iteration counter k which is set to 1. The convergence tolerance ( 0 ) for the coupling constraints is 1 and the parameters (0,1) (e.g., 0.4) and 1 . The augmented Lagrangian multipliers are updated at every iteration as shown in Figure 2.4 while the quadratic penalty parameters are updated only if the improvement of the current 24 iteration is not large enough. The ALO-DQA method alternates between solving an optimization planning problem (4) and solving optimization scheduling sub-problems (5). The solution of each problem is used to linearize the non-separable terms and the algorithm terminates when the consistency constraints (g) have met the pre-defined tolerance or when the given iteration limit is reached. In the next section, three numerical examples are solved using the ALO-DQA method. Initialize, k = 1 0 , 0 0 initial multiplier penalty parameter Solve Planning problem (4) f pl* k , k , k p ,t p ,t P s , Inv s Solving Scheduling problems (5) in parallel f sc*1,1 k , k , k f sc*1,2 k , k , k p ,t f sc* P ,T k , k , k p ,t PP s , II s Loop Converged? g = [Inv-II P-PP] T g ? ? No Update parameters sp ,t sp ,t Ps p ,t - PPs p ,t sp ,t sp ,t Invsp ,t - II sp ,t If g k g k 1 , Set k = k + 1 Yes Feasible solution f * 25 Figure 2.4 Augmented Lagrangian-DQA decomposition algorithm 2.5. Numerical Examples The proposed multiproduct and multisite production facility model is applied to the two examples of supply chain with a planning horizon of 3 months (90 days) and scheduling horizon of 1 production shift (8 hours). The full-scale integrated planning and scheduling problem corresponds to a mixed integer linear programming (MILP) problem while in the ALO-DQA method, the planning problem is quadratic programming (QP) problem and scheduling sub-problems are mixed integer quadratic programming (MIQP) problem. The multisite models were implemented using GAMS 23.6 (2010) and solved using CPLEX 12.2 on 2.53 GHz Precision T7500 Tower Workstation with 6 GB RAM. The scheduling sub-problems (MIQP) in the ALO-DQA method are solved in parallel for each planning period and each production site thus improving the efficiency of the algorithm. In all the examples studied in our work, limited storage capacity for final products and intermediate materials is considered. Example 1: A small example that has 3 production sites serving 3 markets is studied. Each production site contains a multiproduct, multitask batch process plant that produces two products, P1 and P2. (Kondili et al., 1993) The state and task representation (STN) of the production plant is given in Figure 2.5 and the data for the example are given in Appendix Chapter 2. The process parameters, production and inventory costs, and shipping and backorder costs are given in Table A2-1, Table A2-2, and Table A2-3, respectively. Continuous time scheduling problem is solved using 6 event points and 8 hour time horizon. The daily demand data for the example 1 is given in Figure A2-1 for planning horizon of 90 days. 26 Figure 2.5 Production facility state and task network (STN) representation (Example 1) Table 2.1 Model statistics of full-space integrated problem Example 1 Example 2 Example 3 Period Binary Cont. (T) Binary Cont. Const. var. var. 5 720 4006 10 1440 15 Binary Cont. var. var. Const. Const. var. var. 10465 720 5011 10654 1440 8851 21178 8011 20935 1440 10021 21319 2880 17701 42373 2160 12016 31405 2160 15031 31984 4320 26551 63568 30 4320 24031 62815 4320 30061 63979 8640 53101 127153 45 6480 36046 94225 6480 45091 95974 12960 79651 190738 60 8640 48061 125635 8640 60121 127969 17280 106201 254323 90 12960 72091 188455 12960 90181 191959 25920 159301 381493 The full-scale model statistics for example 1 is shown in Table 2.1 and results are given in Table 2.2 for time periods 5 to 90. As the time periods increases, the problem becomes difficult to solve to optimality as observed by the optimality gap (%) in Table 2.2 for the full-space model. The performance and quality of the full-scale model to that of the ALO-DQA is compared in Table 27 2.2. From the Figure 2.6, it can be observed that the augmented Lagrangian algorithm converges to a feasible solution of the original problem as the norm value of the coupling constraints (||g||) converges to zero. The quality of the feasible solution (f*) obtained using the ALO-DQA method may be inferior to the full-scale model since the ALO-DQA strategy solves an approximation version of the relaxation problem. The key information that the integrated model solution provides is production, shipping, sales, inventory, and backorder profiles. The profiles for example 1 obtained using the ALO-DQA method are shown in Figure 2.7 to Figure 2.10 for 30 planning periods. The production profiles for production sites (S1, S2, and S3) are shown in Figure 2.7. The transportation profile of products (P1 and P2) from production site (S1, S2, and S3) to market place (M1, M2, and M3) is given in Figure 2.8, Figure 2.9, and Figure 2.10. Note that as shown in Figure 2.8, Figure 2.9, and Figure 2.10, the production sites 1, 2, and 3 mainly satisfy the demand of markets 1, 2, and 3, respectively. These transportation profiles are expected based on the shipping cost. The total sale of products (P1 and P2) at markets (M1, M2, and M3) is given in Figure 2.11. The variable inventory holding cost is highest at production site 2 and lowest at site 3 and the model solution gives an inventory profiles (Figure 2.12) that has highest holdup at site 3 and lowest holdup at site 2. As expected, the advantage of the proposed decomposition approach is shown for bigger problems. So for larger number of time periods (T=90 periods), better solutions were obtained using the ALO-DQA method than by the full-scale model. Table 2.2 Computational results for example 1 ALO-DQA method Full-space model 0 1.0, 2.0 T CPU Iter. f* sec 5 3600 58073 CPU Gap (%) 0.63 k sec 10 34.68 f* λg + σ||g||2 ||g|| 59500 0.62 0.22 28 10 3600 119227 1.91 9 53.53 122903 34.80 0.66 15 3600 194567 2.07 12 101.97 199983 -157.42 0.89 30 3600 390528 2.27 12 183.21 403208 -17.96 0.85 45 3600 592649 2.44 12 292.25 608059 -166.23 0.76 60 7200 791399 3.28 13 436.03 803867 -128.27 0.59 90 7200 1225376 6.93 15 760.50 1203194 -324.10 0.92 Figure 2.6 Solution procedure of the ALO-DQA method (Example 1 and 90 planning periods) 29 Figure 2.7 Production profiles of products (P1 and P2) at production sites (S1, S2, and S3) obtained using the ALO-DQA method (Example 1 and 30 planning periods) 30 Figure 2.8 Shipment profiles of products (P1 and P2) for production site 1 to markets (M1, M2, and M3) obtained using the ALO-DQA method (Example 1 and 30 planning periods) Figure 2.9 Shipment profiles of products (P1 and P2) for production site 2 to markets (M1, M2, and M3) obtained using the ALO-DQA method (Example 1 and 30 planning periods) 31 Figure 2.10 Shipment profiles of products (P1 and P2) for production site 3 to markets (M1, M2, and M3) obtained using the ALO-DQA method (Example 1 and 30 planning periods) Figure 2.11 Sales profiles of products (P1 and P2) for markets (M1, M2, and M3) obtained using the ALO-DQA method (Example 1 and 30 planning periods) 32 Figure 2.12 Inventory profiles of products (P1 and P2) at production sites (S1, S2, and S3) obtained using the ALO-DQA method (Example 1 and 30 planning periods) Example 2. In example 2, we consider a network of 3 production sites producing 4 different products (P3-P6) and serving 3 global markets. All 3 production sites have batch facilities whose STN network is shown in Figure 2.13. (Kondili, 1987) This batch facility produces 4 products (P3, P4, P5, and P6) through 8 tasks from three feeds and there are 6 intermediates state in the system. The full-scale model statistics for this example are shown in Table 2.1 and results are shown in Table 2.3. The full-scale problem is much easier to solve compared with example 1, even though this example is bigger than example 1. Thus, the production recipe, and the parameters relating to production, capacity, demand, and costs have significant effect on the performance of the full-scale model. To further improve the integrated model performance, we applied the ALO-DQA method and the results are shown in Table 2.3. The solution convergence 33 profile for planning period 90 is shown in Figure 2.14. The performance of the ALO-DQA method depends on the choice of the initial values of Lagrange multipliers, penalty parameters, and other algorithm parameters. By selecting appropriate values of these parameters, we can improve on the quality of the feasible solution obtained by the ALO-DQA. The results with set of parameters 0 0, 0 0, 0 0.20, 1.2, 0.4 are shown in the Table 2.3. Significant CPU time savings is reported when the ALO-DQA method is used compared to the integrated fullscale model. Figure 2.13 Production facility state and task network (STN) representation (Example 2) Table 2.3 Computational results for example 2 ALO-DQA method Full-space model 0 0.2, 1.2 T CPU Gap Iter f* sec λg + CPU sec (%) k f* ||g|| σ||g||2 34 5 52.15 249097 0.00 36 49.37 250052 99.31 0.56 10 3600 502983 0.07 36 66.85 504687 14.18 0.65 15 3600 756527 0.25 40 112.49 762006 -211.75 0.89 30 3600 1381017 0.50 40 209.29 1390602 -216.91 0.95 45 3600 1964466 1.05 40 262.49 1968542 -234.33 0.98 60 3600 2594945 0.75 41 499.90 2608491 -1.59 0.70 90 3600 4009291 1.04 41 709.48 4024795 24.81 0.75 Figure 2.14 Solution procedure of the ALO-DQA method (Example 2, T = 90 periods) Example 3. In example 3, we consider a network of 6 production sites producing 6 different products (P1-P6) and serving 9 global markets. Of the 6 production sites, 3 batch production facilities have the network shown in Figure 2.5 and produce 2 products (P1 and P2) and 3 35 production sites have the batch facilities whose STN network is shown in Figure 2.13 and produce 4 products (P3, P4, P5, and P6) through multipurpose units. Table 2.4 Computational results for example 3 ALO-DQA method Full-space model 0 1.0, 2.0 T CPU Iter f* CPU Gap (%) sec k sec f* λg + σ||g||2 ||g|| 5 3600 141129 4.11 11 71.11 151551 24.28 0.99 10 3600 276230 4.40 13 144.97 295756 609.56 0.88 15 3600 408436 4.97 12 192.53 435078 200.49 0.58 30 3600 784535 6.05 13 378.51 822924 -31.96 0.91 45 3600 1259638 12.97 16 749.96 1233187 -905.25 0.80 60 3600 8123074 82.11 16 984.92 1629968 -718.61 0.76 90 3600 8445397 74.11 15 1263.73 2437213 -1271.17 0.85 36 Figure 2.15 Solution procedure of the ALO-DQA method (Example 3, T = 90 periods) The model statistics of example 3 are shown in Table 2.1. As expected, the complexity of the integrated full-scale model increases as the planning horizon increases and this is reflected in the solution time and relative gap (%) given in Table 2.4. The progress of the solution procedure of the ALO-DQA method for 90 time periods is shown in Figure 2.15. Table 2.4 shows the solutions of the integrated full-scale problem and the ALO-DQA decomposition method. Similar to results of examples 1 and 2, significant computational savings are observed when decomposition is applied compared to the full-scale model. Furthermore, for example 3, the ALO-DQA method is able to provide a better solution than the one reported by the full-scale model for planning periods of T=45, 60, and 90. 37 2.6. Summary This work addresses the problem of integrated planning and scheduling for multisite, multiproduct and multipurpose batch plants using the augmented Lagrangian method. The integrated multisite model is proposed by extending single site formulation of(Z. Li & Ierapetritou, 2010a). The shipping costs from the production sites to distribution markets are taken into account explicitly in the integrated problem. Given the fixed demand forecast the model optimizes the production, inventory, transportation, and backorder costs. Temporal decomposition scheme was developed to address the large scale model resulting from multiperiod planning and scheduling problem. The augmented Lagrangian relaxation with diagonal approximation method allowed solution of the scheduling optimization problems into parallel. Three example problems solved to illustrate the advantages of applying the augmented Lagrangian decomposition scheme. With the proposed decomposition method, faster solution times were realized. Nomenclature Indices i task j units m distribution market n event point p production site s material state t planning period Sets I tasks 38 Ip task that can be performed at site p I jp tasks that can be performed in unit j at site p J units Jp units that are located at site p J ip units that can perform task i at site p M distribution markets N event points Ps sites that can produced final product s PS production sites S material states S mf products that can be sold at market m S fp products that can be produced at site p T planning periods Parameters d sp , m unit transport cost of material s from site p to market m Demsm ,t demand of product s at market m for period t FixCostip fixed production cost of task i at site p hsp holding cost of product s at production site p stcapsp available maximum storage capacity for state s at site p u sm backorder cost of product s at distribution market m max vimin , j , p , vi , j , p minimum and maximum capacity of unit j when processing task i at site p VarCostip unit variable cost of task i at site p 39 ip, j , i p, j constant, variable term of processing time of task i in unit j at site p sc,,ip , sp,i, p proportion of state s consumed, produced by task i respectively at site p Variables bip, j,t,n amount of material processed by task i in unit j at event point n at site p during period t Dsp ,m,t transportation of product s from site p to market m at period t Invsp ,t inventory level of state s at the end of the planning period t for site p stinsp,t initial inventory for state s in planning period t Tf i ,pj,,tn finish time of task i in unit j at event point n in site p during period t Tsip, ,jt,n start time of task i in unit j at event point n in site p during period t U sm ,t backorder of product s at market m in planning period t wvip, ,jt,n binary variable, task i active in unit j at event point n at site s during period t 40 Chapter 3 3. Centralized – decentralized optimization for refinery scheduling This chapter presents a novel decomposition strategy for solving large scale refinery scheduling problems. Instead of formulating one huge and unsolvable MILP or MINLP for centralized problem, we propose a general decomposition scheme that generates smaller sub-systems that can be solved to global optimality. The original problem is decomposed at intermediate storage tanks such that inlet and outlet stream of the tank belong to the different sub-systems. Following the decomposition, each decentralized problem is solved to optimality and the solution to the original problem is obtained by integrating the optimal schedule of each sub-systems. Different case studies of refinery scheduling are presented to illustrate the applicability and effectiveness of the decentralized strategy. The conditions under which these two types of optimization strategies (centralized and decentralized) give the same optimal result are discussed. 3.1. Introduction Production scheduling defines which products should be produced and which products should be consumed in each time instant over a given small time horizon; hence, it defines which run-mode to use and when to perform changeovers in order to meet the market needs and satisfy the demand. Large-scale scheduling problems arise frequently in oil refineries where the main objective is to assign sequence of tasks to processing units within certain time frame such that demand of each product is satisfied before its due date. As the scale of the production problem increases, the mathematical complexity of the corresponding scheduling problem increases exponentially. Decomposition of the initial system into subsystems which are easier to be solved, is a natural way to deal with this type of optimization problems. There are relatively few papers that have addressed planning and scheduling problems using centralized and decentralized optimization strategies providing a comparison of these two approaches. (Kelly & Zyngier, 2008)presented a procedure to find a suitable way to decompose 41 large decision-making problems and compared different decentralized approaches using hierarchical decomposition heuristics. The focus of their work was to find globally feasible solutions to large decentralized and distributed decision-making problems when a centralized approach is not possible. (G. K. Saharidis et al., 2006; G. K. D. Saharidis, Kouikoglou, et al., 2009)studied the problem of production planning in deterministic and stochastic environments and compared centralized and decentralized optimization for an enterprise consisting of two production plants in series producing many different outputs with subcontracting options. (Chen & Chen, 2005)studied a joint replenishment arrangement with a two-echelon supply chain with one supplier and one buyer, facing a deterministic demand and selling a number of products in the marketplace. They proposed both centralized and decentralized decision policies to analyze the interplay and investigated the joint effects of two-echelon coordination and multi-product replenishment on the reduction of total costs. The cost differences between these policies show that the centralized policy significantly outperforms the decentralized policy. (Gnoni et al., 2003)present a case study from the automotive industry dealing with the lot sizing and scheduling decisions in a multi-site manufacturing system. They use a hybrid approach which combines mixed-integer linear programming model and simulation to test local and global production strategies. Their results show that local optimization strategy allows a cost reduction of about 19% compared to the reference actual annual production plan, whereas the global strategy leads to a further cost reduction of 3.5% and a better overall economic performance.(I. Harjunkoski & Grossmann, 2001) presented a decomposition scheme for solving large scheduling problems for steel production which splits the original problem into sub-systems using the special features of steel making. Their proposed approach cannot guarantee global optimality, but comparison with theoretical estimates indicates that the method produces solutions within 1-3% of the global optimum. (Bassett et al., 1996)presented resource decomposition method to reduce problem complexity by dividing the scheduling problem into subsections based on its process recipes. They showed that the overall solution time using resource decomposition is significantly lower 42 than the time needed to solve the global problem. However, their proposed resource decomposition method did not involve any feedback mechanism to incorporate “raw material” availability between sub sections. In this work, the problem of refinery scheduling optimization is addressed with centralized and decentralized decision making process. The chapter is organized as follows. Section 3.2 describes the type of problem studied in this chapter and presents a real case study provided by Honeywell Hi-Spec Solutions. Section 3.3 defines the mathematical formulation of the problem, whereas the decomposition approach is presented in section 3.4. Section 3.5 presents comparative results for centralized and decentralized optimization of the system. Finally section 3.6 draws conclusions. 3.2. Problem Definition In general there are two decision levels in refinery process operations - the planning and the scheduling level. The planning level determines the volume of raw materials needed for the upcoming months (typically 12 months), and the type of final products and the estimated quantities to be ordered, depending on demand forecasts. After determining the yearly plan in the second level we have to determine the optimal production scheduling. The scheduling level determines the detailed schedule of each production unit and CDU for a shorter period (typically 10 days) by taking into account the operational constraints of the system under study. Once the plan is known (the quantities and the types of final products ordered as well as the arrival of raw materials), managers must schedule the production of any unit based on the objective which usually is minimization of the overall makespan or maximization of the total profit. Refinery system considered here is composed of pipelines, a series of tanks to store the crude oil (and prepare the different mixtures), production units and tanks to store the raw materials and the intermediate and final products (see Figure 1). All the crude distillation units are considered continuous processes and it is assumed that unlimited supply of the raw material is 43 available to system. The crude distillation unit produces different products according to the recipes. There are two types of operating scenario for the storage tanks: scenario 1 where material cannot flow out of the tank when material is flowing into the tank at any time interval, that is loading and unloading cannot happen simultaneously (due to security reasons) or scenario 2 where loading and unloading can happen simultaneously. Figure 3.1 represents the system under study which corresponds to a really case of Honeywell Hi-Spec problem. In this system the production starts from cracking units and proceed to diesel blender unit to produce home heating oil (Red Dye diesel) and automotive diesel (Carb diesel and EPA diesel). Cracking unit, 4CU, processes Alaskan North Slope (ANS) crude oil which is stored in raw material storage tanks ANS1 and ANS2, whereas cracking unit 2 (2CU) processes San Joaquin Valley (SJV) crude oil. SJV crude oil is supplied to 2CU via pipeline. Figure 3.1 State and unit representation of Honeywell Hi-Spec problem The products of cracking units are then processed further downstream by vacuum distillation tower unit and diesel high pressure desulfurization (HDS) unit. The coker unit converts vacuum resid into light and heavy gasoil and produces coke as residual product. The fluid catalyzed high pressure desulfurization (FCC HDS) unit, FCC, Isomax unit produce products that are needed for diesel blender unit. The FCC unit also produces by- product FCC gas. The diesel blender blends HDS diesel, hydro diesel, and light cycle oil (LCO) to produce three different final products. The diesel blender sends final products to final product storage tanks. The byproduct FCC gas and 44 residual product Coke is not stored but supplied to the market via pipeline. The system employs four storage tanks to store intermediate products, vacuum resid, diesel, light gasoil, and heavy gasoil. In the system studied in this chapter, the long term plan is assumed to be given and the objective is to define the optimal production scheduling. In such a case the key information available for the managers is firstly the proportion of material produced or consumed at each production units. These recipes are assumed fixed to maintain the model‟s linearity. The managers also know the minimum and maximum flow-rates for each production unit and the minimum and maximum inventory capacities for each storage tank. The different types of material that can be stored in each storage tank are known as well as the demand of final products at the end of time horizon. The objective is to determine the minimum total makespan of production defining the optimal values of the following variables: 1) Starting and finishing times of task taking place at each production unit; 2) Amount and type of material being produced or consumed at each time in a production unit; 3) Amount and type of material stored at each time in each tank. There are seven groups of constraints which guarantee the operational conditions of the system. These are allocation constraints, production and storage capacity constraints, material balance constraints for units and storage tanks, demand constraints, and time sequence constraints. The mathematical formulation for the scheduling problem is presented in the following section. 3.3. Mathematical Formulation In this section a mathematical model is presented for the scheduling of the refinery system presented in section 3.2 where the objective is to minimize the overall makespan. The developed mathematical formulation uses continuous time representation since this leads to reduced number of decision variables and constraints compared to the discrete time representation and also since due to the continuous operating mode, continuous time representation can provide more accurate 45 results. (Ierapetritou et al., 1999; Jia et al., 2003) The mathematical formulation proposed in this chapter has the following main assumptions: 1) the change over time between different tasks at each unit is negligible; and 2) the processes are running at steady state. The proposed model presented in this section minimizes the overall makespan of the refinery production and involves allocation, capacity, storage, material balance, demand, and time sequence constraints. The integer and continuous decision variables used in the developed model give rise to a mixed integer linear programming (MILP) problem. For the minimization of the makespan, four main operating rules have to be followed. The first one is to satisfy the constraint that at most one task can take place in one production unit at one time interval and that at most one material can be stored in one storage tank at one time interval (constraint 1). The second one is to satisfy the material balance constraints (constraints 710). The third one guarantees the demand satisfaction (constraint 11). Finally the forth one guarantees the correct time sequence of the tasks given that a continuous time representation is used (constraints 12-35). Allocation Constraints: Constraints (1) express that if a task (i) starts at event point (n), then it must be performed in one of the suitable units (j). It also satisfies the rule that a unit can physically perform only one task at any given time. wv(i, j, n) 1, j J , n N iI ( j ) (1) Capacity Constraints: Constraints (2) enforce the requirement that material processed by unit (j) performing task (i) at any point (n) is bounded by the maximum and minimum rates of production. The maximum and minimum production rates multiply by the duration of task (i) performed at unit (j) give the maximum and minimum material being processed by unit (j) correspondingly. 46 Rmin (i, j ) Tf (i, j, n) - Ts(i, j, n) b(i, j, n) Rijmax Tf (i, j, n) - Ts(i, j, n) , i I , j J (i ), n N (2) Storage Constraints: Variable in( j, jst, n) is equal to 1 if there is flow of material from production unit (j) to storage tank (jst) at event point (n); otherwise it is equal to 0. Variable out ( jst , j, n) is equal to 1 if material is flowing from storage (jst) to unit (j) at event point (n), otherwise it is equal to 0. Equations (3) and (4) are capacity constraints for storage tank. Constraints (3) state that if there is material inflow to tank (jst) at interval (n) then total amount of material inflow to the tank should not exceed the maximum storage capacity limit. Similarly, constraints (4) state that if there is outflow from tank (jst) at interval (n) then the total amount of material flowing out of tank should not exceed the storage limit at event point (n). infow( j, jst , n) V max ( jst ) in( j, jst, n), j J , jst JSTprodst ( j), n N (3) outflow( jst , j, n) V max ( jst ) out ( jst , j, n), j J , jst JSTstprod ( j ), n N (4) Constraints (5) and (6) represent the requirement that the material in tank (jst) should not exceed the capacity limit V max ( jst ) of this storage tank at any event point (n). st ( jst , n 1) inf low( j, jst , n) inflow1( jst , n) V max ( jst ), jst Jst ( s), n N jJprodst ( jst ) stin( jst ) inflow( j , jst , n) inflow1( jst , n) V max ( jst ), jst Jst (s ), n 0 jJprodst ( jst ) (5) (6) Material Balance Constraints for Operating Unit: Constraints (7) represent the requirement that the production of a unit should be equal to the sum of the amount of flows entering its subsequent storage tanks and reactors, and the delivery to the market. iI ( j ) p ( s, i) b(i, j, n) jstJSTprodst ( j ) Jst ( s ) inflow( j, jst , n) unitflow( s, j, j , n) j Jseq ( j ) Junitc ( s ) outflow2( s, j, n), s S , j J , n N (7) 47 Similarly, constraints (8) represent that the consumption at a unit is equal to the sum of the amount of streams coming from preceding storage tanks, previous units, and stream coming from supply. C ( s, i ) b(i, j , n) inflow2( s, j, n) iI ( j ) jstJSTstprod ( j ) Jst ( s ) unitflow( s, j , j , n), outflow( jst , j, n) j J , s S , n N (8) j 'Jseq j Junitps Material Balance Constraints for Storage Tank: Similar to material balance constraints for units, the material balance constraints (9) and (10) for the storage tanks state that the inventory of a storage tank at one event point is equal to that at previous event point adjusted by the input and output stream amount. St ( jst , n) St ( jst , n 1) inflow( j , jst , n) in flow1( jst , n) jJprodst ( jst ) outflow( jst , j , n) outflow1( jst , n), jst Jst , n N (9) jJstprod ( jst ) St ( jst , n) Stin( jst ) inflow( j , jst , n) inflow1( jst , n) outflow( jst , j, n) outflow1( jst , n), jst Jst , n N jJprodst ( jst ) (10) jJstprod ( jst ) Demand Constraints: Demand for each final product d(s) must be satisfied in centralized problem and also in decentralized problem. Constraints (11) state that production units must at least produce enough material to satisfy the demand by the end of the time horizon. outflow1( jst , n) outflow2(s, j, n) d (s), ( jst , n ) jst Jst ( s), j J , n N , s S ( j ,n) (11) Duration Constraints: Time Sequence Constraints for Each Unit: Constraints (12) to (14) express that if task (i) starts at event point (n+1), then it must start after the end of the same task happening at event point (n) in the same unit (j). If task (i) takes place at unit (j) at event point (n) then wv(i,j,n)=1, and Ts(i,j,n+1) must be greater than or equal to Ts(i,j,n). If wv(i,j,n)=0 then the constraint in equation (12) is relaxed and constraints in equations (13) and (14) enforce the sequencing of tasks. 48 Ts(i, j, n 1) Tf (i, j, n) UH (1 wv(i, j, n)), i I , j J (i), n N (12) Ts(i, j, n 1) Ts(i, j, n), i I , j J (i), n N (13) Tf (i, j, n 1) Tf (i, j, n), i I , j J (i), n N (14) Constraints (15) represent the rule that if task (i’) should happen at event point (n) in unit (j) then task (i) must start at event point (n+1) after the end of task (i’) at event point (n). Ts(i, j, n 1) Tf (i, j, n) UH (1 wv(i, j, n)), j J (i), i I ( j ), i I ( j ), i i , n N (15) Constraints (16) to (19) represent that two consecutive productions, where unit (j) consumes the material produce by unit (j’), with no storage in between, happen at the same time. If wv(i,j,n)=1 and wv(i’,j’,n)=1 then Ts(i,j,n)=Ts(i’,j’,n) and Tf(i,j,n)=Tf(i’,j’,n). If either wv(i,j,n)=0 or wv(i’,j’,n)=0, then the constraints are relaxed. Ts(i, j , n) Ts(i , j , n) UH (2 wv(i, j, n) wv(i , j , n)), j J , j Jseq( j ), i I ( j ), i I ( j ), n N (16) Ts (i, j , n) Ts (i , j , n) UH (2 wv(i, j , n) wv(i , j , n)), j J , j Jseq( j ), i I ( j ), i I ( j ), n N (17) Tf (i, j , n) Tf (i , j , n) UH (2 wv(i, j, n) wv(i , j , n)), j J , j Jseq ( j ), i I ( j ), i I ( j ), n N (18) Tf (i, j , n) Tf (i , j , n) UH (2 wv(i, j , n) wv(i , j , n)), j J , j Jseq ( j ), i I ( j ), i I ( j ), n N (19) Time Sequence Constraints Connecting Unit and Storage Tank: Constraints (20) to (23) state that production and storage occur at the same time. If unit (j) produces the material that is stored in tank (jst) then start and finishing time of production task at unit (j) and inflow to the storage tank must be same. Ts (i, j , n) Tss ( j , jst , n) UH (2 wv(i, j , n) in( j , jst , n)), jst Jst, j Jprodst ( jst ), i I ( j ), n N Ts(i, j , n) Tss( j , jst , n) UH (2 wv(i, j, n) in( j, jst , n)), jst Jst , j Jprodst ( jst ), i I ( j ), n N (20) (21) 49 Tf (i, j , n) Tsf ( j , jst , n) UH (2 wv(i, j , n) in( j , jst , n)), jst Jst, j Jprodst ( jst ), i I ( j ), n N (22) Tf (i, j , n) Tsf ( j , jst , n) UH (2 wv(i, j , n) in( j , jst , n)), jst Jst, j Jprodst ( jst ), i I ( j ), n N (23) Constraints are given by Equations (24) to (27) state that storage and production happen at the same time. If tank (jst) stores the material that is consumed in the unit (j), then outflow from storage tank and production take place at the same time. Ts(i, j , n) Tss( jst , j, n) UH (2 wv(i, j , n) out ( jst , j, n)), jst Jst, j Jstprod ( jst ), i I ( j ), n N Ts(i, j , n) Tss( jst , j, n) UH (2 wv(i, j , n) out ( jst , j , n)), jst Jst, j Jstprod ( jst ), i I ( j ), n N Tf (i, j , n) Tsf ( jst , j , n) UH (2 wv(i, j , n) out ( jst , j , n)), jst Jst, j Jstprod ( jst ), i I ( j ), n N Tf (i, j , n) Tsf ( jst , j, n) UH (2 wv(i, j , n) out ( jst , j , n)), jst Jst, j Jstprod ( jst ), i I ( j ), n N (24) (25) (26) (27) Sequence Constraints for Storage Tank: Constraints are given in Eqns. (28) to (35) connect input and output flow happening at event (n) to the next event point (n+1) for storage tanks. Input starts at event point (n+1) after output finishes at event point (n) and output flow happens at event point (n+1) after the end of input flow happening at event point (n). Binary variable in(j,jst,n) equal to 1 when material is flowing into the tank (jst) from unit (j) at event point (n), otherwise its zero. Similarly out(jst,j,n) equal to 1 when material is flowing from the tank to the unit at event point (n), otherwise its zero. Constraints (28), (31), (34), and (35) are active when the binary variables are equal to 1, whereas when the binary variables are equal to 0, the sequence constraints are enforced by Eqns. (29) , (30), (32), and (33). Tss( j, jst, n 1) Tsf ( j, jst, n) UH (1 in( j, jst , n)), Tss( j, jst , n 1) Tss( j, jst , n), jst Jst , j Jprodst ( jst ), n N jst Jst, j Jprodst ( jst ), n N (28) (29) 50 Tsf ( j, jst , n 1) Tsf ( j, jst , n), jst Jst, j Jprodst ( jst ), n N Tss( jst , j, n 1) Tsf ( jst , j, n) UH (1 out ( jst , j, n)), jst Jst , j Jstprod ( jst), n N (30) (31) Tss( jst , j, n 1) Tss( jst, j, n), jst Jst, j Jstprod ( jst ), n N (32) Tss( jst , j, n 1) Tsf ( jst, j, n), jst Jst, j Jstprod ( jst ), n N (33) Tss ( j , jst , n 1) Tsf ( jst , j , n) UH (1 out ( jst , j, n)), jst Jst , j Jstprod ( jst ), j Jprodst ( jst ), n N (34) Tss ( jst , j , n 1) Tsf ( j , jst , n) UH (1 in( j , jst , n)), jst Jst , j Jstprod ( jst ), j Jprodst ( jst ), n N (35) As we mentioned in section 3.2, there are two types of operating scenario for storage tanks and for each one of them, we have the following additional constraints. Scenario 1: Simultaneous Loading and Unloading Not Allowed Scenario 1 represents that material cannot flow in and out of the storage tank at the same time at any event point (n). This operation rule is used in many refineries for security reasons. Constraints (36) represent that output flow from storage tank (jst) starts after input ends at any event point (n). Tsf ( j , jst , n) UH (1 in( j , jst , n)) Tss ( jst , j , n) UH (1 out ( jst , j , n)), jst Jst , j Jprodst ( jst ), j Jstprod ( jst ), n N (36) Scenario 2: Simultaneous Loading and Unloading Allowed The operation rule for scenario 2 is that material can flow in and out of tank at the same time at any time interval (n). This assumption is very common in many refineries for intermediate storage tanks. Constrains (37) to (40) are enforced when both variables in(j,jst,n) and out(jst,j,n) are equal to 1. When the constraints are enforced, the starting and finishing times of loading and unloading events are equal. When either in(j,jst,n) or out(jst,j,n) is equal to zero, the constraints are relaxed. 51 Tss ( j , jst , n) UH (1 in( j , jst , n)) Tss ( jst , j , n) UH (1 out ( jst , j , n)), jst Jst , j Jprodst ( jst ), j Jstprod ( jst ), n N (37) Tss ( j , jst , n) UH (1 in( j , jst , n)) Tss ( jst , j , n) UH (1 out ( jst , j , n)), jst Jst , j Jprodst ( jst ), j Jstprod ( jst ), n N (38) Tsf ( j , jst , n) UH (1 in( j , jst , n)) Tsf ( jst , j , n) UH (1 out ( jst , j , n)), jst Jst , j Jprodst ( jst ), j Jstprod ( jst ), n N (39) Tsf ( j , jst , n) UH (1 in( j , jst , n)) Tsf ( jst , j , n) UH (1 out ( jst , j , n)), jst Jst , j Jprodst ( jst ), j Jstprod ( jst ), n N (40) Makespan Constraints: Constraints (41) to (46) state that starting and finishing time of any task is always less than equal to the makespan (H). Ts(i, j, n) H , i I , j J (i), n N (41) Tf (i, j, n) H , i I , j J (i), n N (42) Tss( j, jst, n) H , jst Jst, j Jprodst ( jst ), n N (43) Tsf ( j, jst , n) H , jst Jst , j Jprodst ( jst ), n N (44) Tss( jst, j, n) H , jst Jst , j Jstprod ( jst ), n N (45) Tsf ( jst , j, n) H , jst Jst , j Jstprod ( jst ), n N (46) Objective Function: Finally, equation (47) defines the objective function of the problem which is the minimization of makespan. The most common motivation for optimizing the process using minimization of makespan as objective function is to improve customer services by accurately predicting order delivery dates. z min( H ) (47) 3.4. Solution Approach The CPU time for the solution of the overall model presented in the previous section is usually high due to model size. In order to reduce the CPU resolution time we developed a structural 52 decomposition strategy which decomposes the problem into a number of smaller and thus easier to solve subsystems. The developed structural decomposition approach and the additional constraints presented in this section guarantee that the solution obtained by the decentralized optimization will be feasible for the centralized system and is exactly the same. Structural Decomposition Approach Generally scheduling problems are large scale problems and are difficult to be solved to optimality. As the scale of the production increases, the mathematical complexity of the developed model increases, and the CPU time that is required for the solution of the corresponding problem increases too. Decomposition is a natural way to dealing with large scale problems. There are two types of decomposition: the structural decomposition of the system under study and the mathematical decomposition such as Benders decomposition (Benders, 1962), Lagrangian relaxation (Minoux, 1986) etc. Note that mathematical decomposition can be applied after the application of the structural decomposition if it is applicable. The decomposition strategy proposed here decomposes the refinery scheduling problem presented in section 3.2 spatially. To obtain the optimal solution in decentralized optimization approach, each sub-system is solved to optimality and these optimal results are used to obtain the optimal solution for the entire problem. In our proposed decomposition rule, we split the system in such a way so that a minimum amount of information is shared between the sub-problems. This means splitting the problem at intermediate storage tanks such that the inflow and outflow streams of the tank belong to different sub-systems. The decomposition starts with the final products or product storage tanks, and continues to include the reactors/units that are connected to them and stops when the storage tanks are reached. The products, intermediate products, units and storage tanks are part of the sub-system 1. Then following the input stream of each storage tank, the same procedure is used to determine the next sub-system. If input and output stream of the tank are included at the same local problem then the storage tank also belongs to that local problem. 53 Figure 3.2 Intermediate storage tank connecting two sub-systems When the problem is decomposed at intermediate storage tanks, storage tanks become a connecting point between two sub-systems. The amount and type of material flowing out of the connecting intermediate storage tank at any time interval (n) becomes demand for the preceding sub-system (k+1) at corresponding time interval (see Figure 3.2). After decomposing the centralized system, the individual sub-systems are treated as independent scheduling problems and solved to optimality using the mathematical formulation described in section 3.3. It should be also noticed that the operating rules for the decentralized system are the same as those required for the centralized problem. In general the local optimization of sub-system k gives minimum information to the sub-system k+1 which optimizes its schedule with the restrictions regarding the demand of the intermediates obtained by subsystem k. In Figure 3.3 we present the decomposition of the system under study after the application of the developed decomposition rule. The system is split in two sub-systems where sub-system 1 produces all of the final products and one by-product. The sub-system 1 includes 5 production unit, 7 final product storage tanks, and 3 raw material tanks, Raw material tanks in sub-system 1 are defined as intermediate tanks in centralized system. The sub-system 2 includes 4 production units, 1 intermediate tank, 2 raw material tanks and it produces 4 final products. Except Coke, all other final products in sub-system 2 are defined as intermediate products in centralized system. 54 Figure 3.3 Decomposition of the refinery problem under consideration In the following section we present all the additional constraints used in the decentralized approach in order to guarantee that the obtained solution by the sub-system 1 could be realized by the sub-system 2. Feasibility Constraints for Decentralized Model The sub-systems obtained using the decomposition rule presented in the previous section, have all the constraints presented in the basic model but in addition to that the k+1 sub-system has to satisfy the demand of final products produced by this sub-system and also the demand of intermediate products needed by sub-system k. The demand constraints for intermediate final products for sub-system k+1 are given by equation (48). outflow2(s, j, n) r (s, n), s S , j Junitp( s, k 1), n N j (48) Scenario 1: Simultaneous Loading and Unloading Not Allowed When production units in sub-system k+1 supply material to storage tanks located in sub-system k, in order to obtain globally feasible solution, the following capacity constraints are added to sub-system k+1. Constraint in equation (49) is for time interval n=0; sum of the material supplied to storage tank (jst) in sub-system k and initial amount present in the storage tank must be within tank capacity limit. Whereas equations (50) and (51) represents capacity constraints for event point n=1 and n=2 respectively. 55 outflow2(s, j, n) stin( jst ) V max ( jst ), s S , jst Jst ( s, k ), j Junitp ( s, k 1), n 0 j (49) 1 outflow2(s, j, n) stin( jst ) r (0) V j max s n0 ( jst ), (50) s S , jst Jst ( s, k ), j Junitp( s, k 1), n N 2 1 outflow2(s, j, n) stin( jst ) r(s,n) V j n0 max ( jst ), n0 (51) s S , jst Jst ( s, k ), j Junitp( s, k 1), n N Constraints (52) and (53) represent lot sizing constraints for sub-system k+1. The demand of intermediate final product s at event point n is adjusted by the amount present in the storage tank after the demand is satisfied at previous event point (n-1). This adjusted demand is then used in demand constraints for intermediate final products. r ( s,1) outflow2( s, j , 0) stin( jst ) r ( s, 0) r ( s,1), j s S , j Junitp (s , k 1), jst Jst (s , k ) (52) 1 1 r ( s, 2) outflow2( s, j, n) stin( jst ) r ( s, n) r ( s, 2), n 0 j n 0 s S , j Junitp( s, k 1), jst Jst ( s, k ), n N (53) The optimal time horizon of global problem is obtained by combining the optimal schedules of sub-systems at each point (n) such that the material balance constraints are satisfied for connecting intermediate storage tanks. Since sub-system k+1 satisfies the demand of subsystem k, sub-system k+1 will happen before the sub-system k. Scenario 2: Simultaneous Loading and Unloading Allowed To implement the constraint of simultaneous loading and unloading to intermediate connecting storage tanks between any two sub-systems k and k+1, the following constraints are added to subsystem k. inflow1( jst , n) (s) Tsf ( jst , j, n) Tss( jst , j, n) , s S , jst Jst (s, k ), j Junitc( s), n N (54) k 1 k 1 k (s) Pmax (s) if Pmax (s) Cmax (s), s S (55) 56 k k+1 β(s) Ckmax (s) if Cmax (s) Pmax (s), s S (56) k 1 where Pmax (s) is the maximum rate of production of intermediate final product (s) in sub-system k k 1 k k+1 and Cmax (s) is the maximum rate of consumption of (s) in sub-system k. Pmax (s) and Cmax ( s) can be calculated before the start of the optimization process based on the configuration of subsystems. β(s) takes the value of either the maximum rate of production or maximum rate of consumption as given by constraints (55) and (56). Constraint (54) determines the value of material inflow to the storage tank at event point (n). When β(s) takes the value of maximum rate of consumption, constraint (57) is added to the sub-system k+1. outflow2(s, j, n) (s) (Tf (i, j, n) Ts(i, j, n)), s S , j Junitp ( s, k 1), i I ( j ), n N j (57) Constraints (57) state that the total amount of material (s) which is an intermediate final product, produced in sub-system k+1 is equal to the maximum rate of consumption in sub-system k times duration of production in sub-system k+1. inflow1( jst , n) iter outflow2( s, j , n), s S , jst Jst ( s, k ), j Junitp( s, k 1), n N j (58) where iter is a binary variable. After solving sub-systems k and k+1 to optimality respectively, if the optimal makespan results in a time horizon of sub-system k+1 to be greater than sub-system k, then iter = 1. When iter = 1, we resolve sub-system k to optimality with constraint (58) active in the model. Eq. (58) specifies the amount of material flowing into the connecting intermediate storage tank from sub-system k+1 at each event point (n). Since, loading and unloading can happen at the same time, the time horizon of each subsystem will be the same and the global optimal solution can be obtained by combining the optimal schedule of each sub-system. 3.5. Numerical Results The refinery production scheduling case study presented here is based on realistic data provided by Honeywell Hi-Spec Solutions. The data for the problem studied here are presented in 57 Appendix Chapter 3. Different demand cases for final products, Carb diesel, EPA diesel, and Red Dye diesel; and residual products, Coke and FCC gas, are studied. The actual values of the products‟ demands are given in Table A3-1. For all computations in this chapter GAMS/CPLEX 7.0 is used to solve the resulting MILP formulations. The optimal solution is obtained with 1e-6 integrality gap using a Pentium(R) 4 processor at 3.40Hz and with 1.99GB memory. The two scenarios presented in section 3.3, with and without the assumption of simultaneously loading and unloading of tank are examined. Centralized and decentralized optimization is applied using the decomposition approach presented in section 3.4. In the following tables four different examples are presented in order to illustrate the advantage of decentralized optimization using the decomposition approaches. Four examples represent lower to high demands for the system that need to be satisfied within available time horizon of 240 hours. Following the mathematical model presented is section 3.3, the 1st sub-system as shown in Figure 3.3 is solved to minimize the makespan and meet the demand for the final products. Based on the optimal solution of sub-system 1, sub-system 2 is solved to optimality such that it satisfies the demand required by sub-system 1. In the end the solutions of the two sub-systems are combined to obtain the solution of the entire problem. As shown in Table 3.1 and Table 3.2, after the application of the decomposition strategy the size of subsystems is significantly reduced (ex. scenario 1 from 1081 to 749) giving rise to a small increase in the number of constraints. This happens because the decision variables associated with connecting the two sub-systems in centralized problem become demand data in decentralized approach resulting to additional constraint to the first sub-system. Table 3.1 Characteristics of mathematical formulation using Scenario 1 Total Continuous Binary Total number number of Variables Variables Constraints of constraints variables 58 Centralized 985 96 Sub System 1 484 54 1081 1425 921 749 Sub System 2 187 1425 1686 24 765 Table 3.2 Characteristics of mathematical formulation using Scenario 2 Total Continuous Binary Total number number of Variables Constraints Variables of constraints variables Centralized 985 96 Sub System 1 484 54 1081 1488 930 743 Sub System 2 181 24 1425 1485 555 To obtain the global optimal solution for scenario 1, the optimal schedules of each subsystem are combined at each event point n such that the material balance and storage capacity constraints for intermediate connecting tanks are satisfied, whereas the global optimal schedule in scenario 2 is obtained by superimposing the optimal solution of each sub-system. As shown in Table 3.3 and Table 3.4 for both scenarios, the centralized and decentralized optimizations give exactly the same optimal makespan for all examples. As explained in the following section, subsystem 1 has exactly the same solution in centralized and decentralized optimization which gives rise to the same optimal solution (centralized and decentralized) for sub-system 2. The objective function in centralized and decentralized strategy is minimization of makespan. In order to spend minimum time producing material (s), it is required to operate all the units in the system in such a way that they will produce all materials needed at a maximum production rate that is feasible for that particular system. The maximum possible production rate ( Pmax (s) ) for each product (s) can be determined based on the data given in Table A3-2, Table 59 A3-3. Storage tank capacity data are given in Table A3-4. In sub-system 2, only one unit (j) produces the specific material (s), which means that Pmax (s) is defined only by the highest feasible production rate of the unit (j). This highest feasible production rate of unit (j) is calculated by taking under consideration the units that come before and after unit (j) (without any storage in between). Thus, the highest feasible production rate for unit (j) depends only on the parameters and the configuration of the system therefore, it is constant in both centralized and decentralized approach for all event points (n). We can then conclude that the total time spent in order to produce (s) is the same in centralized and decentralized system and this gives rise to identical makespan for sub-system 1. Subsequently, given that the makespan for sub-system 1 is the same in centralized and decentralized system, the sub-system 2 takes the same demand in centralized and decentralized optimization because the final product demand requirements are the same in both optimization approaches. Since, sub-system 2 has the same configuration, same data, and same constraints in centralized and decentralized system, then sub-system 2 has the same optimal solution (may be with different task assignments) concerning the makespan in both optimization approaches. As shown in Table 3.3 and Table 3.4, the CPU time required to find the optimal makespan is reduced significantly in decentralized approach compared to centralized approach. This is mainly due to the reduction in size and complexity of the system going from centralized system to sub-systems. Table 3.3 Scenario 1 Centralized vs. Decentralized Centralized System Ex. CPU Time (s) Objective Value (hr) Decentralized System Total Local CPU- Objective time(s) value(hr) CPU- Global Time(s) Objective(hr) 60 1 2 3 4 318.672 256.187 2192.515 589.984 Sub 1 0.969 Sub 2 2.313 Sub 1 1.141 235.890 97.000 3.282 163.352 235.890 138.89 72.00 4.062 Sub 2 2.921 Sub 1 1.063 79.328 163.352 91.352 36.960 4.047 Sub 2 2.984 Sub 1 1.093 26.149 79.328 42.368 10.272 3.311 Sub 2 2.218 26.149 15.878 Table 3.4 Scenario 2 Centralized vs. Decentralized Centralized System Ex. CPU Time (s) Decentralized System Objective Value (hr) 2 3 4 4.125 5.844 5.562 7.562 Local CPU- Objective time(s) value(hr) Global Time(s) Sub 1 1 Total CPU- Objective(hr) 1.031 235.000 235.000 1.672 Sub 2 0.641 Sub 1 1.047 215.000 235.000 235.000 215.000 1.813 Sub 2 0.766 Sub 1 1.031 191.000 2.343 Sub 2 1.312 Sub 1 1.359 97.000 215.000 215.000 191.000 97.000 2.547 Sub 2 1.188 191.000 97.000 97.000 61 Figure 3.4 Gantt chart of the operation schedule for example 1, centralized system, scenario 1. For scenario 1, the CPU time needed to solve the centralized system is in the order of 100 seconds whereas the decentralized system is solved within 5 CPU seconds. In scenario 2, decentralized solution approach also show improvement compared to centralized system. The CPU time needed to solve the problem is cut by half in decentralized system compared to centralized system. For scenario 1, the production schedule obtained by decentralized approach is different than that obtained by centralized approach as shown in Figure 3.4 and Figure 3.5 for example 1. This difference in production schedule is obtained because in decentralized system, an optimal solution is obtained by integrating the schedules of each sub-system at each event point. Gantt charts for storage tanks are given in Figure A3-1 to Figure A3-10 for centralized and decentralized systems and for scenario 2. 62 Figure 3.5 Gantt chart of the operation schedule for Case 4, decentralized system, scenario 1. 3.6. Summary In this chapter, a structure decomposition strategy and formulation is presented for short-term scheduling of refinery operations. It is shown that the decentralized system model results in fewer constraints and fewer continuous and binary variables compared to centralized system. The chapter presents a problem where both optimization strategies result in the same optimal makespan with the advantage of having significant reduction in the computational time for decentralized system compared to that of centralized system. For decisions making process in refinery where demands always need to be satisfy and determination of the makespan is the main concern, proposed approach provides a decomposition scheme that provides global optimal solution in decentralized system. Furthermore, the proposed decomposition approach is suitable for scheduling of production unit operations because they need to satisfy blend components demand required by blend scheduling operations in time. Nomenclature Indices j Production units jst Storage tanks 63 i Tasks n Event points s States k kth sub-system in decentralized system Sets J Production units Jst Storage tanks S States N Event point within the time horizon J (i) Units which are suitable for performing task i I ( j) Tasks which can be performed in unit j Iseq(i) i produces state s that will be consumed by task i Jstprod ( jst ) Units that consume material s stored in tank jst Jprodst ( jst ) Units that produce material s stored in tank jst Junitp(s) Units that can produce material s Junitp(s, k ) Units in sub-system k that can produce material s Junitc(s) Units that consume material s Junitc(s, k ) Units in sub-system k that can consume material s Jseq( j ) Units that follow unit j (no storage in between) Jst (s) Tanks that can store material s Jst (s, k ) Tanks in sub-system k that can store material s JSTprodst ( j ) Tanks that follow unit j JSTstprod ( j ) Tanks that are followed by unit j Parameters 64 Rmin (i, j ) Minimum rate of material processed by task i required to start production unit j Rmax (i, j ) Maximum rate of material processed by task i in unit j Pmax (s) The possible maximum rate of production of material (s) k Pmax ( s) The maximum rate of production of intermediate final product s in sub-system k k Cmax ( s) The maximum rate of consumption of intermediate final product s in sub-system k V max ( jst ) Maximum available storage capacity of storage tank jst p (s, i) Proportion of state s produced by task i, p (s, i) 0 c (s, i) Proportion of state s consumed by task i, c (s, i) 0 d ( s) Demand of the final product s at the end of the time horizon r (s, n) Demand of intermediate state s at event point n r ( s, n) Adjusted demand of intermediate state s at event point n stin( jst ) Amount of state s that is present at the beginning of the time horizon UH Available time horizon Variables wv(i, j, n) Binary variable that assign the starting of task i in unit j at event point n iter Binary variable that assign the number of iterations between subsystems b(i, j, n) Amount of material undertaking task i in unit j at event point n st ( jst , n) Amount of state s present in storage tank jst at event point n 65 inflow1( jst , n) Flow of raw material to storage tank jst event point n outflow1( jst , n) Flow of final product from storage tank jst at event point n inflow2(s, j, n) Flow of raw material s to production unit j at point n outflow2(s, j, n) flow of product material s from unit j at point n inflow( j, jst , n) Flow of material from unit j to storage tank jst event point n outflow( jst , j, n) Flow of material from storage tank jst to unit j at point n in( j, jst , n) Binary variable that assign the starting of material flow into storage tank jst from unit j at point n out ( jst , j, n) Binary variable that assign the starting of material flow out of storage tank jst to unit j at point n unitflow(s, j, j , n) Flow of state s from unit j to consecutive unit j’ for consumption at point n st ( jst , n) Amount of material in tank jst at event point n Ts(i, j, n) Time that task i starts in unit j at event point n Tf (i, j, n) Time that task i finishes in unit j at event point n Tss( j, jst , n) Time that material starts to flow from unit j to storage tank jst Tsf ( j, jst , n) Time that material finishes to flow from unit j to tank jst at event point n Tss( jst , j, n) Time that material starts to flow from tank jst to unit j at event point n Tsf ( jst , j, n) Time that material finishes to flow from tank jst to unit j at event point n H Time horizon 66 Chapter 4 4. Scheduling of a large-scale oil-refinery operations Refineries are increasingly concerned with improving the scheduling of their operations in order to achieve better economic performances by minimizing quality, quantity, and logistics give away. In this chapter, we present a comprehensive integrated optimization model based on continuous-time formulation for the scheduling problem of production units and end-product blending problem. The model incorporates quantity, quality, and logistics decisions related to real-life refinery operations. These involve minimum run-length requirements, fill-draw-delay, one-flow out of blender, sequence dependent switchovers, maximum heel quantity, and downgrading of better quality product to lower quality. The logistics giveaways in our work are associated with obtaining a feasible solution while minimizing violations of sequence dependent switchovers and maximum heel quantity restrictions. A set of valid inequalities are proposed that improves the computational performance of the model significantly. The formulation is used to address realistic case studies where feasible solutions are obtained in reasonable computational time. 4.1. Introduction The oil-refinery production operation is one of the most complex chemical industries, which involves many different and complicated processes with various connections. Over the last 20 years, it has grown increasingly more complex due to tighter competition, stricter environment regulations, and lower margin profits. Main objective of the oil-refineries is to transform crudeoil into gasoline, diesel, jet fuel, and other middle distillate hydrocarbon products that can be used as either feedstock or energy source in chemical process industry. The short-term scheduling is a critical aspect in this large and complex production process. The refinery operations scheduling problem involves decisions that are related to quantity, quality and logistics. Quantity decisions 67 include lot-sizes for raw material, intermediate and product tanks inventories, amount of material moving between production units and storage tanks, etc, while quality decisions deal with obtaining finished products that meet specific quality requirements. Logistics constraints include policies and procedures for production operations which deal with allocating resources to operations, sequencing or ordering of different modes of operations, and determining the durations of operations. To remain competitive in dynamic global marketplace, the oil-refineries are increasingly concerned with improving the scheduling of their operations in order to achieve better economic performance by minimizing product quality violations, incidence of high quality product giveaway, and logistics giveaway. Oil-refinery operations can be classified into three sub-operations based on the structure of the refinery configuration as shown in Figure 4.1. (Jia et al., 2003) These sub-operations are (1) crude-oil unloading and mixing, (2) production unit operations, and (3) finished products blending and distribution. Sub-operation 1 involves the crude-oil unloading, blending, and inventory control, Sub-operation 2 includes production unit scheduling, and Sub-operation 3 consists of finished product blending and lifting. Typically, suboperations scheduling optimization problems are addressed separately since centralized optimization approach gives rise to an incomprehensive large scale mixed integer non-linear programming (MINLP) models. Review of refinery scheduling problem has been presented in work of N. K. Shah et al. (2011). The crude-oil unloading and mixing scheduling problem has been studied by Pedro M. Castro and Grossmann (2014), Lee et al. (1996), Jia et al. (2003) , Kolodziej et al. (2013), Reddy et al. (2004a), Reddy et al. (2004b), G. K. D. Saharidis, Minoux, et al. (2009), and G. K. D. Saharidis et al. (2010). The complex crude-oil blend scheduling optimization problem is decomposed into logistics and quality sub-problems by Kelly and Mann (2003a, 2003b). They utilized successive linear programming (SLP) to solve the quality sub-problem. Once the crude prepared mixture is prepared by crude oil loading and unloading operations, it is charged to CDUs for distillation. The distillation cuts from CDU are then sent to other production units for 68 fractionation and reaction to produce blend components for finished products. The most common refinery includes catalytic, hydro, and thermal cracking units to convert heavy hydrocarbons into light hydrocarbons. They also include other process units like continuous catalytic reforming, hydro treating, and hydro desulfurization units. The production unit operations scheduling problem is characterized by continuous processes, intermediate component storage, and recycle stream. Figure 4.1 Graphic overview of a standard refinery system The production units scheduling problem based on continuous time representation is proposed by Joly et al. (2002) and Jia and Ierapetritou (2004) where they applied spatial decomposition to solve the scheduling problem. Gary et al. (2007) is a comprehensive reference for refinery production unit operations. The finished products blend scheduling has received significant attention in the literature. The purpose of blend scheduling optimization problem is to find the best way of mixing different semi-finished products that have been rectified during various refinery processes with some additives so as to produce final products that meet quality specifications and demand while minimizing cost. The gasoline blending operation is highly nonlinear and gives rise to a mixed integer non-linear programing scheduling model. Blend recipe optimization problems considering nonlinear properties relations have been studied extensively. 69 (Castillo & Mahalec, 2014a, 2014b; Cuiwen et al., 2013; Glismann & Gruhn, 2001; J. Li & Karimi, 2011; J. Li et al., 2010; Mendez et al., 2006; Zhao et al., 2008) Glismann and Gruhn (2001) proposed a decomposition technique based on first solving the nonlinear (NLP) quality optimization model and then solving a MILP model to optimize temporal and resource decisions. J. Li and Karimi (2011) developed a slot-based MILP formulation for an integrated treatment of recipe, specifications, blending, and storage considering real-life features such as multipurpose product tanks, parallel nonidentical blenders, minimum run lengths, changeovers, piecewise constant profiles for blend component qualities and feed rates, etc. To reduce the computational complexity that arises from non-linear model, many have used constant component properties. (Jia & Ierapetritou, 2003; Joly et al., 2002) Kelly (2006) emphasized the importance of logistics details in refinery blending and delivery problem and proposed a decomposition of the blend scheduling problem into two sub-problems, logistics and quality. The logistics sub-problem considers only the quantity and logistics related variables and the problem constraints whereas the quality sub-problem considers product specifications, quantity constraints and bounds. Their work is based on discrete time representation and their formulation includes many logistics details such as minimum run-length, sequence dependent changeovers, and fill-draw-delay. They observed that incorporating logistics details into scheduling problem can yield substantial improvements in efficiency and productivity. A MILP optimization model based on continuous time representation, using unit specific event points, and fixed blend recipe was developed by Jia and Ierapetritou (2003). They modeled multipurpose product tanks, but they do not include certain features such as sequence dependent switchovers constraints, fill-draw-delay at product tank, and one flow out of blender. An iterative procedure is proposed by Mendez et al. (2006) to deal with variable recipe and nonlinear properties for different grades of products by replacing MINLP with sequential MILP formulations. They enforced prepared blend recipe whenever it is possible. In their model formulation they did not consider multipurpose product tanks, fill-drawdelay, and minimum run-length requirement. There is extensive literature on the refinery blending 70 problem using nonlinear optimization tools. (Adhya et al., 1999; Audet et al., 2004; Gounaris et al., 2009) With more product grades, stricter specifications, new government regulations, and fewer feasible blends, refineries face increasing challenges to maintain, let alone increase, profitability. What is needed is a comprehensive model for the entire refinery scheduling problem and for various refinery configurations that addresses quality, quantity, and logistics issues in a unified, yet flexible approach. The comprehensive model is complex, hard to build and solve and there is sparse work in literature in this area. Moro et al. (1998) proposed a planning model for refinery diesel production where the emphasis is on blending relations. Pinto et al. (2000) proposed a planning and scheduling model for refinery production and distribution operations. Their formulation is based on discretization of time and the model includes features such as sequence dependent transition cost of products within oil pipeline. C. Luo and Rong ; Chunpeng Luo and Rong (2007) developed a two-tiered decision-making hierarchical scheduling model for overall refinery. The upper level optimization model is based on discrete time formulation and it is used to determine sequencing and timing of operations modes and to decide the quantities of materials produced/consumed at each operations mode. The upper decisions level uses aggregated tanks storage capacity whereas the lower level uses heuristics to obtain a detailed schedule. They consider multipurpose product tanks via an iterative procedure that allows to readjust aggregated tank capacity at the lower level by changing multipurpose tank service mode and then to recalculate corresponding optimal solution at the upper level. The logistics details are ensured through heuristics at the lower level. There are also several commercial tools available for refinery scheduling such as Aspen Petroleum Scheduler, Aspen Refinery Multi-Blend Optimizer, Honeywell‟s Production Scheduler, and Honeywell‟s Blend. Honeywell‟s Production Scheduler has a logistics solver to optimize logistics and quantity problems and a quality solver to solve quantity and quality problems. Most of the models utilized in these commercial tools are based on discrete time formulation. 71 In this work, we present a comprehensive integrated optimization model for the production units and end-product blend scheduling problem that incorporates quantity, quality and logistics decisions related to real-life refinery operations. The model is driven by the shipment plan and accounts for the tradeoffs between costs of keeping inventory and changing run-modes. The goals of refinery operations scheduling are to maximize the profit and performance while minimizing the penalties subject to quantity, quality, and logistics giveaways and nonattainment. (Kelly, 2003a) The outline of the chapter is as follow. Section 4.2 presents the problem definition, section 4.3 presents the mathematical formulation which is applied to a realistic case study example to illustrate the applicability of proposed model to large scale model in section 4.4, and the chapter concludes with section 4.5. 4.2. Problem Definition The production scheduling determines the detailed schedule of each production unit and each demand order unloading for a short time period (typically 10 days - 1 month) by taking into account the operational constraints of the plant. The schedule defines which products should be produced and which materials should be consumed in each time interval over a given small time horizon; hence, it defines which run-mode to use and when to perform changeovers in order to meet the market needs satisfying the demand and product specifications. Large-scale scheduling problems arise frequently in oil refineries where the main objective is to assign sequence of tasks to processing units within certain time frame such that the demand of each product is satisfied before its due date while minimizing the cost or maximizing the total profit. The refinery production system considered here is composed of raw material storage tanks, production units, blending units, intermediate tanks, and final product tanks. Each production unit is defined as a continuous processing element that transforms the input streams into several products according to the variable production recipe. For reasons of operating flexibility and cost effectiveness, refinery unit operations can generate a range of intermediate 72 streams which are blended into finished products. For simplicity, we separate the final products into two groups: a) products that are stored in tanks and b) products that are not stored in tanks but are supplied to the market directly from production units. In this work, we limit the use of term “demand order” only for the first group of products and each demand order corresponds to only one kind of product since each multiproduct order can be decomposed into several single product orders. The characteristics of the problem considered in this chapter are given in detail in the next subsection. Problem Characteristics The key information available for the refinery include the following: 1) Maximum and minimum proportion of material produced or consumed at each production unit 2) Maximum and minimum production flow-rates for each production unit 3) Maximum and minimum inventory capacities for each storage tank, identity of material type that each tank can service, and initial holdup in each tanks 4) Upper limit on the flows out of finished product tanks 5) Demand orders for group A products and their delivery time windows 6) Total demand for group B products 7) Maximum allowable heel quantity for multipurpose product tanks 8) Minimum run-length for units and maintenance time for multipurpose tanks between runmodes 9) Quality specifications limit on blend product properties 10) Available scheduling time horizon of 10 days The goal of optimization is to determine: 73 1) The sequencing of tasks for production units 2) Each product pool that satisfies demand orders 3) Durations of tasks at production units and duration of unloading tasks from product pools 4) The inventory levels in component and product pools 5) Production rates for units and unloading rate for product pools 6) Composition of material produced and consumed The problem is also restricted by a series of logistics details as follows: 1) At any given time, only one task can take place at production unit 2) Minimum run-length constraint for production units 3) Non-contiguous product order fulfillment for product in group A 4) Blend unit can send product to multiple pool sequentially, not simultaneously 5) Product tanks cannot distribute and receive material at the same time 6) Fill-draw-delay restriction for product pools enforces certain amount of downtime on tanks after product loading event has taken place 7) Multipurpose tanks can store different types of materials over time, but only one type of material at any given time 8) Sequence dependent switch over for multipurpose tanks where higher quality product is stored before lower grade products 9) Maximum heel requirement restriction does not allow product heel to exceed specified maximum heel quantity when the multipurpose tank is switched to a different mode 10) Downgrading of the higher grade product to lower grade product if necessary Although a realistic case study is modeled the following assumptions had to be made: 74 1) Unlimited supply of raw materials 2) Fixed recipe for crude distillation units (CDUs) 3) Constant blend components properties 4) Perfect mixing in the blender 5) Product tank cannot satisfy multiple demand orders simultaneously 6) Each demand order involves only one product 7) For production units, the amount of time required for run-modes change is neglected 8) Changeover times in multipurpose tanks from higher to lower grade product are negligible Before presenting the mathematical model, a case study with realistic data provided by Honeywell Process Solutions (HPS) is presented in the next section. The refinery produces diesel fuels, jet fuel, and components for gasoline production. This case study will be used to illustrate the applicability of the proposed model in section 4.3. 4.2.1. Case study description The production process at Honeywell refinery consists of 2 blender units, 13 other processing units, and 2 non-identical parallel crude distillation units (CDUs) that process two different type of crude oils. The schematic of production system is shown in Figure 4.2. There are two charging tanks for one CDU and a charging pipeline for another CDU. The CDUs concurrently transform crude oil into several distillation cuts. These distillation cuts from the CDUs are then sent to other production units for fractionation and reaction to produce blend components for finished products. The oil-refinery under case study has following production units: Vacuum Tower, Coker conversion unit, Continuous catalytic reforming (CCR) process unit, Isomax, Fluid catalytic cracking (FCC) unit, Penex, De-C5, and Alkylation process unit. Current regulatory requirements to produce ultra-low-sulfur fuels require the use of hydrotreating technology. Thus, 75 the refinery also includes three hydrodesulfurization (HDS) units: Naphtha HDS, Diesel HDS, fluid catalytic cracking HDS (FCC HDS). Figure 4.2 Schematic of the refinery production plant for reference example Honeywell refinery utilizes Jet blender and Diesel blender units to blend components produced by other production units to produce final products. Jet blender unit blends straight run jet, coker jet, and isomax jet streams to produce jet fuel which can be stored in 2 product tanks. Diesel blender unit produces three different grades of fuel: CARB diesel, EPA diesel, and Reddye diesel. Light cycle oil, HDS diesel, and hydrocracked diesel are blended to produce these three grades of diesel products using three different run-modes. There are two dedicated tanks for each grade of diesel products and one multipurpose tank that can service CARB and EPA diesel. There is 6 hours of cleaning or maintenance downtime when the multipurpose tank service switches from lower grade of diesel product to higher grade of product. This cleaning downtime is essential to remove any sulfur contamination present in the tank before low sulfur 76 product is sent for storage. The product tank has 4 hours of down time called fill-draw-delay for certificate of analysis preparation and to let the product settle down and mix before is shipped to the market. 4.3. Mathematical Formulation In this section we present the mathematical formulation for the refinery production scheduling problem based on continuous time representation and the idea of unit specific event points.(M. G. Ierapetritou & C. A. Floudas, 1998; M.G. Ierapetritou & C.A. Floudas, 1998; Ierapetritou et al., 1999) A state-task network (STN) representation introduced by Kondili et al. (1993) is used to describe the refinery operations. The model involves material balance constraints, capacity constraints, demand constraints, quality constraints, logistics constraints, and setup constraints. Material balance constraints connect the amount of material at one event point to next event point, storage and production capacity limit is enforced by capacity constraints, and demand constraints ensure that all the products demand is satisfied while quality constraints ensure product quality specifications. Logistics constraints include all the logistics details presented in the previous section. If a feasible solution that satisfies all the quantity, quality, and logistics constraints cannot be obtained, then it is essential to produce a schedule that can still be implemented in real-life refinery sacrificing model feasibility. In this case we introduced artificial variables to treat any infeasibility present and these variables are subsequently penalized in the objective function to obtain an optimal solution that satisfies as many as possible from the quantity, logistics, and quality constraints by minimizing giveaways. A detailed description of the each variable and parameter used in the model can be found in the nomenclature section. Variable recipe constraints: Upper and lower bounds are forced on the individual components volumetric flow rates processed at each production units. Here Bi , j ,n is a total amount of material processed at unit j performing 77 task i at event point n . Constraint (1a) enforces the bound for the amount produced, whereas constraint (1b) ensures that the amount consumed is restricted by the imposed recipe. sp,i,min Bi , j , n bps,i , j , n sp,i,max Bi , j , n , s S , i I sP , j J i , n N (1a) sc,,min Bi , j , n bcs,i , j , n sc,,max Bi , j , n , i i s S , i I sC , j J i , n N (1b) Furthermore, the amount of material processed is equal to the total amount of material consumed or produced. Therefore, constraints (1a-1b) can be replaced by (2a-2c) and we can eliminate variable Bi , j ,n . Constraint (2c) satisfies material balance at each production unit j , which states that the total amount of material consumed is equal to the total amount of material produced. bp sp,i,min bps ,i , j , n sp,i,max s , i , j , n s Sip bc sc,,min i s , i , j , n bcs ,i , j , n sc,,max i s Sic bp s ,i , j , n bp s , i , j , n , s S , i I sP , j J i , n N s Sip bc s , i , j , n , s S , i I sC , j J i , n N s Sic bc s ,i , j , n , j J , i I j , n N sSic sSip (2a) (2b) (2c) In our work we assume that the crude distillation units have the same lower and upper bounds which means that the distillation cuts are assumed to be known. Material balance constraints for production units: Constraints (3a) connect the material produced at production units to subsequent storage tanks, production units, and end product delivery to market. Constraints (3b) represents that the consumption at a production unit is equal to the amount of material coming from preceding storage tanks, previous units, and raw material supply. bp iI j bc iI j s ,i , j , n s ,i , j , n Kif s , j , k , n Kof s , k , j , n k K jpk JJf s , j , j , n Uof s , j , n , s S , j J sp , n N (3a) JJf s , j , j , n Uif s , j , n , s S , j J sc , n N (3b) c j J seq j Js Ks k K kp j Ks p j J seq j Js Material balance constraints for storage tanks: The material balance constraints for storage tanks are given by equations (4a-4b). The equations state that the inventory of a tank at one event point is equal to that of previous event point 78 adjusted by the input and output streams amount and by taking into account the downgraded products amount. sts , k , n stos , k Kif s , j ,k ,n Rif s ,k ,n jJ kpk std s Sk sts , k , n sts , k , n 1 Kof Kif s , s , k , n s, j ,k ,n std s Sk Rif s , k , n jJ kpk s , j ,k ,n jJ kkp std s Sk s , s , k , n std Kof s Sk s , s , k , n oOs , s S , k K s , n 1 jJ kkp s , s , k , n Lfk ,o ,n s , j ,k ,n (4a) Lf k ,o , n oOs , s S , k K s ,1 n N (4b) The variable std s, s, k , n is defined as the tank heel of material s present at the end of event point n 1 that is downgraded to material s during event point n . 0 if product s ' is downgraded to s in tank k at event point n std s, s , k , n otherwise 0 When there is a changeover from higher grade product to lower grade, the tank heel present in the tank would be transformed into lower grade product without violating any product property specifications of lower grade product. Capacity constraints for production units: Constraint (5) enforces that the material processed by unit j performing task i is bounded by the maximum and minimum rate of production. Constraint (6) gives an upper bound on total amount of material processed at unit j over the entire time horizon. Rimin , j Tf i , j , n Tsi , j , n bp s , i , j , n bp s ,i , j , n s Sip UH Rimax , j wvi , j , n , Rimax i I , j Ji , n N , j Tf i , j , n Tsi , j , n , i I , j J i , n N s Sip (5) (6) Capacity constraints for storage tanks: Constraints (7a-7c) are capacity constraints for storage tanks and they define the binary variables associated with flow in and out of the tanks. 79 Kif s , j , k , n Vkmax ins , j , k , n , j J , k K jpk , n N Kof s , j , k , n Vkmax outs , j ,k ,n , (7a) j J , k K kp j ,n N (7b) Lf s , k ,o, n Vkmax lk ,o,n , s S , k Ks , o Os , n N (7c) Constraints (8a-8c) represent that the material present in the tank should not exceed maximum storage capacity. The multipurpose tanks can store different grades of products and if the higher grade product is present in the tank when it is being serviced for lower grade product, then the high quality product will be downgraded into lower quality. The downgraded products are taken into consideration for storage capacity limit in constraints (8a-8b). stos , k Kif s, j ,k ,n Rif s , k , n jJ kpk sts , k , n 1 Kif s, j ,k ,n Rif s , k , n jJ kpk Kof jJ kkp s,k , j,n Lf oOs std s Sk k ,o, n s , s , k , n std s Sk Vkmax ys , k , n , s , s , k , n std s S k s , s , k , n std s Sk Vkmax ys , k , n , s , s , k , n Vkmax ys , k , n , s S , k K s , n 1 s S , k K s ,1 n N s S , k K s , n N (8a) (8b) (8c) The maximum and minimum unloading (lift) rate for product storage tanks must be bounded as specified by constraint (9). RU kmin Tof k , o, n Tosk , o, n Lf k , o, n RU kmax Tof k , o, n Tosk , o, n , k K p , o O, n N (9) Quality constraints: The final products produced by the blenders should satisfy the quality specifications. These product qualities are assumed to be computed by volumetric average to maintain model linearity. Since the blend components properties are assumed to be constant, linearity of the model is preserved. Constraint (10) guarantees that final product leaving the outlet port of the blender satisfies set product quality range. Here, Psmin and Psmax , p are the upper and lower limit of property p ,p for final blend product s. 80 l Psmin , p bps , i , j , n pg s , p , n iI sp iI sp , s Sic u Ps , p bcs ,i , j , n Psmax s Sb , j J sp , n N , p bps , i , j , n pg s , p , n , iI sp (10) When the final products produced by the blenders cannot meet the quality specifications at event point n, we introduced positive slack variables pgsl , p,n and pgsu, p,n which are penalized in objective function to minimize giveaways. Demand constraints: Demand of each finished product must be satisfied during the entire scheduling horizon. Constraint (11) guarantees that sufficient amount of product will be available to meet the demand. Do, s rs dg ol rg s Lf k ,o , n Uof s , j , n Do, s dgou , s Sb , o Os s S f k ,n j ,n (11) Due to production capacity limitation, sometimes the demand order of finished product cannot be satisfied during the entire scheduling horizon. A feasible solution can be obtained by introducing the positive artificial variables dgol , dgou and rgs in the demand constraint which are penalized in the objective function to minimize quantity giveaway. Logistics Constraints: Allocation constraints: Constraint (12) expresses that if a task i starts at event point n , then it must be performed in one of the suitable units j . It also satisfies the operating detail that a unit can physically perform only one task at any given time. wv iI j i, j ,n 1, j J , n N (12) The requirement that multipurpose storage tanks can store only one type of products at any time is enforced by equation (13). y sK s s,k ,n 1, k K m , n N (13) Minimum run-lengths: The minimum run-length for each task is enforced by constraint (14). Here the minimum run-length ( RLi ) is 15 hours. 81 Tf i , j , n Tsi , j , n RLi wvi , j , n UH 1 wvi , j , n , i I , j I j , n N (14) Similar to minimum run-length constraint, maximum run-length restriction can be imposed if necessary. Loading and unloading constraints: Product tanks cannot load and unload material at the same time and this restriction is enforced by equations (15). Furthermore, we restrict that product tanks can only satisfy one demand order at any given time. in s, j ,k ,n sSk lk , o , n 1, k K , j J kpk , n N (15) oO A one-flow-out restriction for blender unit: A one-flow-out restriction given by equation (16) is required for all operation tasks on a blender to ensure that the product output from the blend unit can only go to one product tank. This restriction is imposed because the refinery has a blend property online controller that is set up to fill a specific product tank by taking into account the tank heel properties. If a blend unit does not comply with this restriction and send the output from the blender to multiple tanks simultaneously and not sequentially, then this miss-operation can result in a significant off-specification of product stocks. ins , j , k , n 1, s Sb , j J sp , n N (16) k K s K jpk Set-up constraints: To model tank set-up, we use the binary variables k ,n which are defined by constraints (17a-17b). 1 if storage tank k becomes active at event point n for first time 0 otherwise k ,n k ,n k ,n y s,k ,n y s,k ,n sSk sSk ys , k , n , k K h , n N , yos ,k 0 (17a) ys , k , n , k K h , n N , yos , k 0 (17b) sSk , n n s Sk , n n s s 82 Similarly for production unit, the setup variables are j ,n and constraints (18a-18b) represent the utilization of unit at event n for the very first time during the production time horizon. 1 if unit j becomes active at event point n for first time 0 otherwise j ,n j , n wvi , j , n iJ i j , n wvi , j , n iJ i wvi , j , n , j J h , n N (18a) wvi , j , n , j J h , n N (18b) iJ i , n n i J i , n n Since, the refinery operates in a continuous mode; we only define the setup variables for identical parallel production units and tanks. Setup variables are penalized in the objective function to minimize the total number of units and storage tanks that are utilized during refinery operation. Changeovers constraints: Different mode of production and storage tasks are specified by different types of product being produced or stored. Changeovers between modes of operations cause disturbances and additional costs. Thus, few changeovers (long sequences of the same mode of operations) are desired. Continuous variables i,i , j ,n and s, s ,k ,n denote changeover of task at production unit j and changeover of service mode at product pool k, respectively. 1 if mode changes from i ' at event point n to i at later event point 0 otherwise i , i , j , n 1 if mode changes from s ' at event point n to s at later event point 0 otherwise s , s , k , n Changeover constraints proposed by Shaik et al. (2009) are used in this work. Constraints (1920c) are thus used to force the changeover variables to 1 if there is a change in operations mode from event point n to any later event point. i,i , j ,n wvi, j ,n , j J m , i I j , i I j , i i, n N (19a) 83 i,i , j , n 1 wvi , j , n wvi , j , n i I j i I j , nn n n wvi , j , n , j J m , i I j , i I j , i i, n N , n n N i ,i , j , n wvi, j , n wvi , j , n 1 i I j , n n n wvi , j , n , j J m , i I j , i I j , i i,, n N , n n N s, s, k , n ys,k ,n , k K m , s Sk , s Sk , s s, n N s , s , k , n 1 ys , k , n y s Sk s , k , n s Sk , n n n n s Sk , n n n (19c) (20a) ys , k , n , k K m , s Sk , s Sk , s s, n N , n n N s , s , k , n ys , k , n ys , k , n 1 (19b) ys , k , n , k K m , s Sk , s Sk , s s,, n N , n n N (20b) (20c) The changeover variable os, s ,k in equation (20d) is active if there is material s’ present in the tank at the beginning of the time horizon and then service is changed over to new material s at event point n=1. yos , k ys,k ,n os , s, k 1, k K m , s Sk , s Sk , s s, yos ,k 1 (20d) Changeovers variables i,i , j ,n , os, s ,k and s, s ,k ,n are penalized in the objective function to minimize the changeovers. Heel requirement: When changeover occurs from higher to lower quality product, the holdup in the tank must be less than equal to the maximum heel quantity specified. The maximum heel requirement can be enforced on std s , s,k ,n as soft constraint using equations (21a-21b) and positive slack variable mhs ,k ,n . The artificial variable mhs ,k ,n is penalized in the objective function to minimize the heel. std s , s , k , n mhs , k , n Vkheel Vkmax 1 os , s , k , k K , s K s , s K s , s s , n 1 (21a) std s , s , k , n 1 mhs , k , n 1 Vkheel Vkmax 1 s , s , k , n , k K , s K s , s K s , s s , n N (21b) 84 Product downgrading: The downgrading of product happens, 1) there is a changeover of service at used multipurpose tanks or 2) it is required to meet lower quality product lifting demand order due to production capacity limitation. The downgrading of product is captured by equations (22a22e). The variable associated with downgrading std s , s,k ,n is zero when the switchover occurs from lower to higher quality product. stds , s, k , n 0, k K m , s Ks , s K s , s s , n N (22a) stds , s, k , n Vkmaxos , s,k , k K m , s Ks , s Ks , s s , n 1 (22b) stds , s, k , n 1 Vkmaxs , s, k , n , k K m , s Ks , s Ks , s s , n N (22c) std s , s, k , n Vkmax 1 os , s , k stos , k std s , s , k , n Vkmax 1 os , s ,k , k K m , s K s , s K s , s s , n 1 std s , s , k , n 1 Vkmax 1 s , s , k , n sts , k , n std s , s , k , n 1 Vkmax 1 s , s , k , n , k K m , s K s , s K s , s s , n N (22d) (22e) Timing Constraints: Sequence Constraints for production units: Finishing time of any task must be greater than the starting time of that task, as represented by constraint (23a). Constraint (23b) expresses that if task i starts at event point n 1 , then it must start after the end of the same task happening at event point n while equation (23c) enforces the time sequence constraint for different tasks happening in the same unit. Tfi , j ,n Tsi , j ,n , i I , j Ji , n N (23a) Tsi , j , n1 Tfi , j , n , i I , j J i , n N , n N (23b) Tsi , j , n 1 Tf i , j , n UH 1 wvi , j , n , j J , i I j , i I j , i i, n N , n N (23c) 85 Constraints (24a-24d) represent that two consecutive productions with no storage in between, happen at the same time because production units operate as continuous processes. Here, unit j consumes the material produced by unit j . Tsi , j , n Tsi , j , n UH 2 wvi , j , n wvi , j , n , j J , j J seq j , i I j , i I j , n N (24a) Tsi , j , n Tsi , j , n UH 2 wvi , j , n wvi , j , n , j J , j J seq j , i I j , i I j , n N (24b) Tf i , j , n Tf i , j , n UH 2 wvi , j , n wvi , j , n , j J , j J seq j , i I j , i I j , n N (24c) Tf i , j , n Tf i , j , n UH 2 wvi , j , n wvi , j , n , j J , j J seq j , i I j , i I j , n N (24d) Sequence Constraints for storage tanks: Finishing time of the inlet, outlet transfer service has to be greater than or equal to the start time of that service. Constraints (25-27) enforce the sequence time requirement for movement transfer task from one event point to next event point for same unit-tank connection. Tsf j , k , n Tss j , k , n , k K , j J kpk , n N (25a) Tss j , k , n 1 Tsf j , k , n , k K , j J kpk , n N , n N (25b) Tsf k , j , n Tssk , j , n , k K , j J kkp , n N (26a) Tssk , j , n 1 Tsf k , j , n , k K , j J kkp , n N , n N (26b) Tofk ,o,n Tosk ,o,n , k K p , o O, n N (27a) Tosk ,o, n 1 Tof k ,o, n , k K p , o O, n N , n N (27b) Start time sequence constraints for tanks receiving/sending material from/to multiple production units is given by constraints (28a) and (28b). Whereas constraints enforcing time sequence requirement for different type of demand orders satisfied by the tank is given equation (28c). 86 Tss j , k , n 1 Tsf j , k , n UH 1 ins , j , k , n , k K , j J kpk , j J kpk , j j , n N , n N (28a) Tssk , j , n 1 Tsf k , j , n UH 1 outk , j , n , k K , j J kkp , j J kkp , j j , n N , n N (28b) Tosk ,o, n 1 Tof k , o, n UH 1 lk ,o , n , k K p , o O, o O, o O, n N , n N (28c) Sequence constraint for material transfer in and out of intermediate tanks happening at the same event point is enforced by equations (29a-29d). Tss j , k , n UH 1 ins , j , k , n Tssk , j , n UH 1 out s , k , j , n , k K , j J kpk , j J kkp , n N (29a) Tss j , k , n UH 1 ins , j , k , n Tssk , j , n UH 1 outs , k , j , n , k K , j J kpk , j J kkp , n N (29b) Tsf j , k , n UH 1 ins , j , k , n Tsf k , j , n UH 1 outs , k , j , n , k K , j J kpk , j J kkp , n N (29c) Tsf j , k , n UH 1 ins , j , k , n Tsf k , j , n UH 1 outs , k , j , n , k K , j J kpk , j J kkp , n N (29d) Constraints (30a-30b) connect material transfer from one event point to next event point. Product tanks cannot simultaneously load and unload material and this restriction is enforced by equations (31a-31b). Variable tak is a fill-draw-delay parameter for tank k . The tank unloading happens anytime after the material flow into the tank is over and fill-draw-delay downtime has elapsed. Product flow into the tank starts after the end of the product unloading. Tss j , k , n 1 Tssk , j , n UH 1 outs , k , j , n , k K , j J kpk , j J kkp , n N , n N (30a) Tssk , j , n 1 Tss j , k , n UH 1 ins , j , k , n , k K , j J kpk , j J kkp , n N (30b) Tosk , o , n 1 Tsf j , k , n UH 1 ins , j , k , n tak ins , j , k , n , s S , k K s , j J kpk , o Os , n N (31a) Tss j , k , n 1 Tof k ,o , n UH 1 lk ,o , n , k K , j J kpk , o O, n N (31b) The tank needs to go through cleaning maintenance to store higher grade product after servicing lower grade product. This maintenance downtime requirement is captured by constraints (31c). 87 Constraints are relaxed if the there is no switchover in the service from lower grade to higher grade product type. Tss j , k , n 1 Tof k , o, n tcleank 2 s, s , k , n 1 , k K , s S k , s S s , s s , j J kpk , o Os , n N (31c) Sequence Constraints for production units and storage tanks: Upstream production and material flow into storage tank happens at the same time, which is imposed by constraints (32a-32d). Tsi , j , n Tss j , k , n UH 2 wvi , j , n ins , j , k , n , s S , k K s , j J kpk , i I j , n N (32a) Tsi , j , n Tss j , k , n UH 2 wvi , j , n ins , j , k , n , s S , k K s , j J kpk , i I j , n N (32b) Tf i , j , n Tsf j , k , n UH 2 wvi , j , n ins , j , k , n , s S , k K s , j J kpk , i I j , n N (32c) Tf i , j , n Tsf j , k , n UH 2 wvi , j , n ins , j , k , n , s S , k K s , j J kpk , i I j , n N (32d) For intermediate storage tanks, downstream production and material flow out of feed tank occur simultaneously. This constraint is imposed by equations (33a-33d). Tsi , j , n Tssk , j , n UH 2 wvi , j , n out s , k , j , n , s S , k K s , j J kkp , i I j , n N (33a) Tsi , j , n Tssk , j , n UH 2 wvi , j , n out s , k , j , n , s S , k K s , j J kkp , i I j , n N (33b) Tf i , j , n Tsf k , j , n UH 2 wvi , j , n out s , k , j , n , s S , k K s , j J kkp , i I j , n N (33c) Tf i , j , n Tsf k , j , n UH 2 wvi , j , n out s , k , j , n , s S , k K s , j J kkp , i I j , n N (33d) All tasks should start and finish before the end of the scheduling time horizon as stated in (34a34d). Tsi , j ,n H , Tfi , j ,n H , Tss j , k , n H , Tsf j , k , n H , i I , j J i , n N j J , k K jpk , n N (34a) (34b) 88 Tssk , j , n H , Tsf k , j ,n H , j J , k K kp j ,n N (34c) Tosk ,o,n H , Tofk ,o,n H , o O, k K p , n N (34d) Intermediate due dates: Intermediate due dates for group A products, which are stored in product pools, are given by constraints (35a-35b). Orders can start unloading anytime after the vessel arrival time and finish unloading anytime before vessel departure time. The due date requirements are enforced as inequality constraints to consider demurrage by using slack variables Tearlyo and Tlateo which are penalized in the objective function. Tosk ,o , n UH 1 lk ,o , n timeso Tearlyo , o O, k K p , n N (35a) Tof k ,o , n UH 1 lk ,o , n timef o Tlateo , o O, k K p , n N (35b) 4.3.1. Valid Inequities Valid inequalities are added in the proposed model to improve the computational efficiency of the proposed model. Constraints (36a-36b) enforce that if the material s is flowing into the tank at event point n , then the binary variable ys,k ,n is 1 and similarly, if the material is flowing out of the tank, then the binary variable is also 1. in s, j ,k ,n jJ kpk ys , k , n , k K , s S k , n N jJ kpk out s, j ,k ,n jJ kkp ys , k , n , k K , s S k , n N jJ kkp (36a) (36b) If the tank is sending or receiving the material from a production unit, then that unit is active and these requirements are represented by equations (37a-37b). in s, j ,k ,n k K jpk out k K kp j k K jpk s,k , j ,n wvi , j , n , j J , s S jp , n N (37a) wvi , j , n , j J , s S cj , n N (37b) iI j I sp k K kp j iI j I sc 89 Since loading and unloading cannot happen at the same event point, if the material is unloaded at event point n 1 then tank is not empty at previous event point n .This is feature of model is captured by equation (38). l k , o , n 1 y sSk o s,k ,n , k K , n N , n N (38) If two units are consecutive without any storage tank between them, then constraint (39) impose the simultaneous operation of these units due to the continuous operation mode. However, this constraint is not imposed on parallel production units that can produce the same type of products. j J seq j wvi , j , n j J seq j / J , i I j h wvi , j , n , j J , j J h , n N / J , iI j h (39) Constraint (40) enforces the material balance constraint in addition to the constraint presented in equation (2c). Kof s , k , j , n JJf s , j , j , n Uif s , j , n p j J seq j Js k K kp j Ks k K jpk Kif s , j , k , n Ks JJf s , j , j , n Uof s , j , n , j J , n N (40) c j J seq j Js Objective Function: The objective function (41) is used to maximize the performance and profit of total production. The refinery performance is represented by the minimization of utilization of units and tanks, all the connection between production units and tanks, start up set-ups, changeovers, and production downgrading. The blending task is significantly improved by reducing the quantities of downgraded products. Logistics and quality give-aways, under and over production, and demurrage are penalized. Profit term includes the costs of feeds and revenue of products. The penalty weights are assigned arbitrary to each term depending on its importance in schedule. The deviation from intermediate due dates is heavily penalized, quality give-aways are penalized the second most, while connection between unit and tank is least heavily penalized. It is favorable that the multipurpose product tanks store different grade products only in a certain order that is 90 allowed by the sequence dependent switchovers constraint. The favorable switchovers are from higher grade of product to lower grade and unfavorable switchovers are from lower grade to higher grade product. Due to contamination issues, unfavorable switchovers are more heavily penalized than favorable. Similar to multipurpose tanks, there is a sequence dependent switchover restriction for multipurpose blend units. The switchover from the run mode that produces better quality product to lower quality product is least penalized than vice versa. z i , jJ i , n ci1, j wvi , j , n s , k K s , n ck2 ys , k , n c outs , k , j , n 4 k, j j , sS cj , k K s , n sSb , p , n c std s , s , k , n c mhs , k , n 11 s,k k K m , sS k , n oO c Tlateo oO cs8, s s , s , k , n k K , sS k , s S k , s s , n 10 k K m sSb , k K s , oOs , n k , sS k , n oO c Rif s , k , n 19 s c ci7,i i ,i , j , n jJ m ,iI j ,i I j ,i i , n u 14 l 15 u c13 s , p pg s , p , n co dg o co dg o 18 o ck6 k , n k K h , n m k K m , sS k , s S k , s s , n c 4j , k ins , j , k , n j , sS jp , k K s , n c j ,n cs8, s os , s , k k K , sS k , s S k , s s m k K p , o , n ck3lk , o, n 5 j j J h , n c sS f 16 s Uif s , j , n 19 s s , jsc , n s , k K s , n sSb , p , n ck9 sts , k , n l c12 s , p pg s , p , n (41) rg s c17 o Tearlyo oO cs20Uof s , j , n sS f , jsp , n cs21 Lf s , k , o , n Note that the different penalty parameters have significant effect on the computational time required to obtain an optimal solution. 4.4. Results on the Case Studies In this section two problems based on the realistic case study presented in section 4.2.1 are solved to optimality and analyzed to show the effectiveness of the proposed model. Each case study includes different set of examples that differ in either demands, intermediate due dates, or initial hold up in the tank. All the problems are solved on a Dell Precision (Intel R XeonTM with CPU 3.20 GHz and 2 GB memory) running on Windows XP using CPLEX 12.1.0/GAMS 23.2. Raw materials and finished product prices are given in Table 4.1 and the penalties parameter used in the objective function are given in Table 4.2. Table 4.1 Price of raw materials and final products 91 Material Price Material Price Material Price ANS crude oil 25 Jet Fuel 70 Pentane 40 SJV crude oil 20 Refinery Gases 50 NC4 40 Carb Diesel 80 Isomerate 60 Alkylate 60 EPA Diesel 60 Reformate 60 FCC Gas 45 Red Dye Diesel 50 Isomax Gasoline 65 Coke 30 Table 4.2 Penalty parameters in objective function Penalty Parameter Value Penalty Parameter Value 1 Ccarbnormal , diesel blender 125 Cs11,k 10 1 CEPAnormal , diesel blender 105 Cs12,p 1000 1 CRe d dyenormal , diesel blender 85 Cs13,p 1300 Ck2 1 Co14 1150 Ck3 1 Co15 500 C 4j , k 4 (multipurpose tank: 6) Cs16 1100 C 5j 1 Co17 1400 Ck6 150 Co18 1250 Ci7,i 40 (Unfavorable: 50) Cs19 price(s) (e q)1 Cs8, s 60 (unfavorable: 70) Cs20 price(s) e1 Cs21 price(s) q1 Ck9 V Ck10 V max 1 k heel k Where, e 3 rs and q 2 ordero , s . s o,s 1 92 4.4.1. Case study 1 The first case study is obtained by deleting three diesel product tanks to reduce the model size and to obtain medium scale case study. The medium scale refinery has only one dedicated tank for each three diesel product and one multipurpose tank that can service CARB and EPA diesel. For this problem, there are no product tanks receiving material from more than one blender unit and no raw material tank supplying material to multiple CDUs. Thus we exclude constraints (28a) and (28b) from the model because the time sequence requirements for intermediate tanks are satisfied by other sequence constraints present in the formulation. Data for different set of examples is given in Table 4.3 and Table 4.4 and results are shown in Table 4.5 and Table 4.6. The initial hold up in Table 4.4 is stated for multipurpose tanks. Demand of group B products has to be met before the end of the scheduling time horizon. In this case study, the scheduling horizon is 72 hours (3 days). The products P1, P2, P3, and P4 correspond to Red dye diesel, EPA diesel, CARB diesel and Jet fuel, respectively. In this case study, we have included 4 product qualities requirements for blend products. The rule of thumb for the smallest number of event points needed to obtain an optimal solution is (d+1) where d represents the total number of diesel products in the demand orders. For example, if only CARB and EPA diesel products are required, then 3 event points should be considered first. For the examples addressed in the case study 1, the aforementioned valid inequalities allowed us to compute medium scale scheduling problems with significantly less computational effort. The first integer solution is obtained within 15 seconds for all the examples studied. The computational performance without and with valid inequalities is presented in Table 4.5 and Table 4.6, respectively. When the valid inequalities are included in the model, the number of variables remains the same, but the number of constraints and nonzero elements increase. Valid inequalities have no effect on quality of the optimal solution; rather their effect is concentrated in significantly reducing the computational effort needed to find the optimal solution. The CPU time required to reach optimal solution is reduced by 90% when the valid inequalities are present 93 versus when they are not included in the model. When example 3 is solved without valid inequalities, the optimal solution is not obtained even after 28 hours, whereas, when it is solved with valid inequalities, the optimal solution is obtained within 1 hour. The size of the model increases as the event point increases and time to obtain optimal solution also increases as observed for examples 4. In example 1, in order to satisfy the demand of order 2 (O2) within its delivery window, higher quality product P1 is downgraded to lower quality product P2. Multipurpose tank inventory data for example 1 is shown in Table 4.7. Product degradation is observed in optimal solution of example 3 in order to satisfy the product demand of order 2. Table 4.3 Group A products demand order data for case study 1 Orders (Product type, amount(kbbl), delivery window, delivery rate(kbbl/h)) Ex. O6 O7 P4 P4 P2 [10,75] [50,150] [25,75] [10,38] [36,48] [54,65] [25,38] [55,72] [63,72] 10 10 10 10 10 10 P3 P2 P1 P1 P4 P4 P2 [5,15] [5,21] [15,50] [10,75] [50,150] [25,75] [10,38] [10,23] [23,30] [36,48] [54,65] [25,38] [55,72] [63,72] O1 O2 O3 O4 O5 P3 P2 P1 P4 [10,100] [50,150] [50,175] [10,150] [58,71] [10,20] [28.5,46.5] [40,70] 3 10 10 3 P3 P2 P1 P4 P4 [37,100] [55,150] [98,175] [112,150] [50,100] [65,72] [55,68] [30,43] [38,50] [59.6,71.5] 10 10 10 10 10 P3 P2 P1 P1 [5,15] [5,21] [15,50] [10,23] [23,30] 10 1 2 3 4 94 10 10 10 10 10 10 10 Table 4.4 Group B products demands data for case study 1 Group B products demand (kbbl) Initial holdup Ex. (product, kbbl) P5 P6 P7 P8 P9 P10 P11 P12 P13 1 P1 - 10 5 20 20 20 5 0 5 5 0 2 - 5 24 65 13 5 0 5 5 0 3 P1 - 7 5 20 15 8 5 3 5 4 9 4 P2 - 7 5 20 15 8 5 3 5 4 9 Table 4.5 Computational performance of case study 1 (without valid inequalities) 0-1 Ex. Cont. n Nonzero Const. Var. Var. Nodes CPU Obj. Time (s) Value Iterations Elements % gap 1 4 260 1683 4200 14841 167899 24414456 12504.70 1125.98 0.00 2 4 268 1719 4311 15245 379243 58903042 35273.88 1643.92 0.00 3 4 284 1793 4542 16024 1000000 161530640 102634.73 1234.92 3.75* 4 4 284 1794 4541 16021 1000000 204485505 121729.67 1193.73 5.01* 4 5 355 2242 5743 20702 564000 218078428 169094.88 1065.40 16.30** CPU Obj. % Time (s) Value gap *Nodes limit reached, ** out of memory Table 4.6 Computational performance of case study 1 (with valid inequalities) Ex 0-1 Cont. n . Nonzero Const. Var. Var. Nodes Iterations Elem 1 4 260 1683 4474 15675 10552 1487406 713.28 1125.98 0.00 2 4 268 1719 4585 16085 20085 2684973 1309.78 1643.92 0.00 3 4 284 1793 4816 16876 56424 7003479 3473.47 1234.24 0.00 95 4 4 284 1794 4815 16873 14922 1810327 983.47 1193.54 0.00 4 5 355 2242 6087 21772 78318 17482935 11616.69 1055.45 0.00 Table 4.7 Material flow in and out of multipurpose tank for case study 1, example 1 Event Points Initial Hold Up Flow Direction (product/amount, kbbl) n1 loading n2 P2 (40 kbbl) P1/10 unloading P2 (50 kbbl) 4.4.2. Case Study 2 In this section the proposed model is applied to Honeywell Hi-Spec refinery problem presented in section 4.2.1. We exclude constraints (28a) and (28b) because of the reasons mentioned in the previous case study. The time horizon considered in this case study is 10 days (240hrs) and there are four different quality restrictions placed on the blend products. The data for different demand orders are shown in Table 4.8 and Table 4.9 and results for these data are shown in Table 4.10. All the examples reach the first integer solution within 20 seconds. In many instances, the optimal solution is reached fast and the rest of the time is spent proving the global optimality which is a typical behavior of mixed integer programming models. In example 1, the optimal solution is obtained with product P4 sulfur limit violations and due date violation (demurrage) of demand order 1. When only 4 event points are used for example 3, the optimal solution of 3866.22 is obtained that does not fully satisfy demand order 7 (O7). However, when 5 event points are used, the optimal solution of 1562.13 is obtained that satisfies all the demand orders within their due dates. In the case of example 5 with 5 event points, the first integer solution of 286630.74 and 99.78% gap is reached within 10 seconds and the first integer solution that does not violate any demand requirements and due date restrictions is obtained within 1600 96 seconds with an objective value of 1741.81 and gap of 58.89 %. Example 6 obtains the first integer solution with objective value of 287161.44 and 99.91% gap in 15 seconds. The solution without any quantity, quality, and demurrage violations is reached within 450 seconds with objective value of 1728.68 and 82.08% gap. A solution without product downgrading is obtained for example 6 when 6 event points are used. As the demand order increases, the event points needed to reach best solution also increases, thus size of the model increases too. Table 4.8 Group A products demand order data for case study 2 Product type, amount(kbbl), delivery window, delivery rate (10 kbbl/h) Orders O1 O2 O3 O4 O5 O6 Ex 1 Ex 2 Ex 3 Ex 4 Ex 5 Ex 6 P3 P3 P3 P3 P1 P1 [50,100] [75,100] [15,50] [15,50] [15,50] [15,50] [28.5,46.5] [110,140] [36,54] [36,54] [36,54] [24,50] P2 P2 P2 P2 P2 P1 [50,150] [50,100] [32,50] [32,50] [32,50] [10,50] [65,89] [63,89] [63,92] [63,92] [63,92] [63.73] P1 P1 P1 P1 P3 P2 [50,200] [75,125] [45,70] [45,70] [45,70] [5,70] [105,130] [26.5,60] [103,119] [103,119] [103,119] [85,109] P1 P1 P1 P1 P1 P3 [50,175] [100,175] [63,98] [63,98] [63,98] [13,98] [216.5,235.5] [215,235.5] [135,148.5] [135,148.5] [135,148.5] [125,138.5] P4 P4 P4 P4 P4 P4 [75,200] [100,200] [50,150] [50,150] [50,150] [20,150] [90,120] [90,120] [100,135] [100,135] [100,135] [100,125] P4 P4 P4 P4 P4 P4 [50,250] [75,250] [15,130] [15,130] [15,130] [25,130] [174,200] [204,230] [200,236.5] [200,236.5] [200,236.5] [200,236.5] 97 O7 P2 P2 P2 P2 P2 P1 [50,100] [50,100] [10,48] [10,48] [10,48] [10,48] [166,211.5] [166,191.5] [166,199.5] [166,199.5] [166,205] [145,175] P3 P3 P1 P2 [10,35] [10,35] [10,35] [10,35] [226,240] [226,240] [226,240] [195,220] O8 P3 O9 [10,35] [230,240] P4 O10 [25,150] [225,240] Table 4.9 Group B products demands data for case study 2 Group B products demand (kbbl) Initial holdup Ex. (product, kbbl) P5 P6 P7 P8 P9 P10 P11 P12 P13 1 - 10 50 50 50 10 10 15 15 0 2 - 5 50 200 50 50 10 40 10 50 3 P2-8 7 43 172 53 40 11 44 15 30 4 - 7 43 172 53 40 11 44 15 30 5 - 7 43 172 53 40 11 44 15 30 6 P1-10 2 13 72 30 10 5 22 5 0 Table 4.10 Computational results for case study 2 (with valid inequalities) Last Integer Solution Var. Optimal Solution Event Ex. Int./Cont. Nodes/ Obj. CPU Gap Nodes/ Obj. CPU (Constraints) Iterations Value Time (%) Iterations Value Time (s) Points 98 (s) Nonzero Elem. 328/1956 103384/ 1 4 (5333) 249153/ 34638.64 5000 0.45 9312323 34638.64 7910.20 1671.86 2449.83 3866.22 12762.77 1562.13 112879.86 1743.87 5150.23 1649.75 3787.66 1511.44 136811.42* 1137.57 234163.58** 997.25 161182.4*** 14282981 18614 328/1957 20760/ 2 4 (5333) 35489/ 1671.86 1800 6.57 3626885 4741152 18622 336/1994 79969/ 3 4 (5431) 207687/ 3866.22 8300 3.64 15494045 23440793 18947 420/2490 419747/ 3 5 (6872) 872647/ 1562.13 100700 1.96 119257871 133443528 24466 336/1993 43412/ 4 4 (5438) 59005/ 1743.87 4650 6.70 8367727 9238856 18992 340/2009 22787/ 5 4 (5518) 55538/ 1649.75 2650 3.16 4473417 6353433 19244 425/2510 97379/ 5 5 (6976) 509585/ 1511.44 21500 44.30 30910982 178283024 24793 445/2597 428485/ 6 5 (7246) 916986/ 1137.57 132500 38.10 160549580 281680876 25694 534/3114 120720/ 6 6 (8778) 375832/ 997.25 54800 52056975 31779 * Out of memory. Optimal solution obtain with 33.13% gap ** Out of memory. Optimal solution obtain with 12.62% gap 40.71 150550479 99 *** Out of memory. Optimal solution obtain with 30.59% gap 4.5. Summary In this chapter, a short-term scheduling model is developed based on continuous time representation for large-scale refineries. The model features logistics decisions such as start-up, minimum run-length, fill-draw-delay, one-flow out of blender, sequence dependent changeovers, maximum heel quantity, and downgrading of product. A set of valid inequalities are proposed that reduces the CPU resolution time by a significant factor for large scale refinery problems. The model with valid inequalities is applied to different examples and was observed that valid inequalities result in up to 90% reduction in CPU performance time compared to model without inequalities. The model is applied to two case studies to illustrate the applicability of the proposed formulations to large scale refinery operations. However, even with inclusion of the valid inequalities in the scheduling model, computational expense required to reach an optimal solution is still considerably high for the real-life oil-refinery applications. Hence, there is a need to employ different decomposition approaches such as mathematical, heuristics, or combination of both heuristics and mathematical decomposition to enable the solution of large-scale problems in a reasonable timeframe and bridge the gap between theory and industrial applicability. Nomenclature Indices i Tasks j Production units k Storage tanks n Event points o Product order p Properties 100 s States Sets Ij Tasks which can be performed in unit j I sp Tasks which can produce material s I sc Tasks which can consume material s J Production units J sc Units that consume material s Ji Units which are suitable for performing task i Jh Units that can produce all the same products as some other unit in the refinery Jm Units which are suitable for performing multiple tasks J kkp Units that consume material s stored in tank k J kpk Units that produce material s stored in tank k J sp Units that can produce material s J seq j Units that follow unit j (no storage in between) K Storage tanks Kh tanks that can store the same products as some other tank in the refinery K kp j Tanks that store material consumed by unit j Km Multipurpose tanks that can store multiple materials Kp Tanks that can store final products K jpk Tanks that store material produced by unit j Ks Tanks that can store material s N Event point within the time horizon O Orders for products that are stored in tanks P Product Properties 101 S States Sb Group A final products, produced by blenders and stored in tanks Sf Group B final products, products that are not stored in tanks Sk Materials that can be stored in tank k Sic Materials that can be consumed by task i S cj Materials that can be consumed by unit j Sip Materials that can be produced by task i S jp Materials that can be produced by unit j Parameters Do, s , Do, s Demand limit requirement for order o and product s that is stored in tank rs Demand of the final product s at the end of the time horizon max Rimin , j / Ri , j Minimum/ maximum rate of material be processed by task i in unit j RLi Minimum run length for task i RU kmin / RU kmax Minimum/maximum rate of product unloading at tank k stos,k Amount of state s that is present at the beginning of the time horizon in k tak Fill draw delay for product tank k UH Available time horizon Vkmax Maximum available storage capacity of storage tank k Vkheel Maximum heel available for storage tank k yos , k 1 if the material s is present at the beginning of the time horizon in k max smin ,i / s ,i Proportion of state s produced/consumed by task i, Variables Binary Variables 102 wvi , j , n Assignment of task i in unit j at event point n ins , j ,k ,n Assigns the material flow of s into storage tank k from unit j at point n Assigns the starting of product flow out of product tank k to satisfy order o at event point lk ,o,n n outs , k , j , n Assigns the material flow of s out of storage tank k into unit j at point n ys , k , n Denotes that material s is stored in tank k at event point n Positive variables bps ,i , j , n Amount of material s produced task i in unit j at event point n bcs ,i , j , n Amount of material s undertaking task i in unit j at event point n dgol Minimum demand quantity give-away term for order o dgou Maximum demand quantity give-away term for order o H Total time horizon used for production tasks JJf s , j , j , n Flow of state s from unit j to consecutive unit j’ for consumption at point n Kif s , j ,k ,n Flow of material s from unit j to storage tank k event point n Kof s , k , j , n Flow of material s from storage tank k to unit j at point n Lfo, k , n Flow of final product for order o from storage tank k at event point n mhs , k , n Maximum heel give-away term for product tanks pg sl , p, n Lower limit giveaway of product quality p pg su, p, n Upper limit giveaway of quality p for product s rgs Minimum demand quantity give-away term for Group B product s Rif s, k , n Flow of raw material to storage tank k event point n sts , k ,n Amount of state s present in storage tank k at event point n std s , s, k , n Amount of state s that is downgraded to state s‟ in storage tank k at event point n 103 Tearlyo Early fulfillment of order o than required Tfi , j ,n Time that task i finishes in unit j at event point n Tlateo Late fulfillment of order o than required Tosk ,o, n Time that material starts to flow from tank k for order o at event point n Tof k ,o, n Time that material finishes to flow from tank k for order o at event point n Tsi , j ,n Time that task i starts in unit j at event point n Tsf j ,k ,n Time that material finishes to flow from unit j to tank k at event point n Tsf k , j , n Time that material finishes to flow from tank k to unit j at event point n Tss j , k , n Time that material starts to flow from unit j to storage tank k Tssk , j , n Time that material starts to flow from tank k to unit j at event point n Uif s , j ,n Flow of raw material s to production unit j at point n Uof s , j , n Flow of product material s from unit j at point n j ,n For unit j, 1 if the unit becomes active for very first time at event point n k , n For tank k, 1 if the tank becomes active for very first time at event point n s , s , k , n os , s, k i , i , j , n Continuous 0-1 variable, 1 if material in tank k switchover service from s at event point n to s‟ at later event point Continuous 0-1 variable, 1 if material in tank k switchover service from s to s‟ Continuous 0-1 variable, 1 if task at unit j changes from i at event point n to i’at later event point. 104 Chapter 5 5. Lagrangian Decomposition Approach for Refinery Operations Scheduling In this chapter, mathematical decomposition based on Lagrangian relaxation is proposed for the scheduling of refinery operations from crude oil processing to the blending and dispatch of finished products. A new algorithm for Lagrangian decomposition (LD) is proposed and applied to realistic large scale refinery scheduling problem to evaluate its efficiency. A novel strategy is presented to formulate restricted relaxed sub-problems based on the solution of the Lagrangian relaxed sub-problems that take into consideration the continuous process characteristic of the refinery. This new restricted algorithm provides tighter lower bound compared to classical Lagrangian decomposition approach. Furthermore, proposed heuristic rules leads to faster improvement in upper bounds. The goal of the mathematical decomposition is to produce better solutions for those integrated scheduling problems that cannot be solved in reasonable computation times. The application of the proposed algorithm results in substantial reduction in CPU solution time, duality gap, and the total number of iterations compared to classical LD. 5.1. Introduction Lagrangian relaxation is often used for NP-hard problems where it is observed that many problems can be considered easy problems made difficult because of presence of complicating constraints. Lagrangian decomposition creates relaxed problem that is relatively easy to solve than original problem by relaxing these complicating constraints and dualizing them in the objective function with multipliers. Lagrangian relaxation provides a lower bound on the global optimal for minimization problem. Relaxed Lagrangian problems can be decomposed into easy to solve sub-problems by fixing Lagrangian multipliers. These multipliers are then typically updated iteratively using a subgradient method (Fisher, 1985) based on the upper bound. Upper bound is usually obtained using heuristics at every iteration and iterations proceeds until duality gap lies 105 within predefined tolerance. The use of Lagrangian relaxation was popularized in the early 1970s (Cornuejols et al., 1977; Fisher, 1973, 1981; Held & Karp, 1971; Held & Karp, 1970) and this approach has wide ranging application such as scheduling (Adhya et al., 1999; Ghaddar et al., 2014; Luh & Hoitomt, 1993; Wu & Ierapetritou, 2003), planning (Graves, 1982; Gupta & Maranas, 1999; Tang & Jiang, 2009), and integrated planning and scheduling (Calfa et al., 2013; Z. Li & Ierapetritou, 2010a; Mouret et al., 2011). Lagrangian relaxation is also used to calculate lower bound in outer approximation algorithm for crude oil scheduling by Karuppiah et al. (2008) and in branch and bound approach by Holmberg and Hellstrand (1998); Holmberg and Yuan (2000). Refinery operations involve crude oil unloading and blending, production unit operations, and finished product blending and delivery as mention in previous chapter. The model proposed in previous chapter is an integrated unified model for production unit operations and blending operations scheduling that can be applied to different type of refinery configurations. These refinery configurations differ based on blending modes. Traditionally, blending is carried out from blend component tanks, normally used for gasoline, while “rundown blending” is normally used for diesel, jet fuel, and kerosene. In rundown blending many, if not all component streams come directly from a process unit with no component storage. Many refineries are forgoing the use component tanks due to economic pressure for reducing on-site storage inventory and safety concerns for reducing inventory of volatile materials. In many instances when blend scheduling optimization problem is investigated, traditional blending operations with components storage tanks are considered. (Castillo & Mahalec, 2014a, 2014b; Cuiwen et al., 2013; Glismann & Gruhn, 2001; J. Li & Karimi, 2011; J. Li et al., 2010; Mendez et al., 2006; Zhao et al., 2008) However, scant attention has been paid to rundown blending operations optimization problem in refinery. These traditional blending scheduling models are not designed to optimize blending operations without component storage tanks. In recent years, a long term swing in demand for 106 diesel is expected to increase by 75% from 2010 to 2040 due to higher fuel efficiency of diesel engines and preferential tax treatment of diesel "The Outlook for Energy: A View to 2040" 2013). Furthermore, demand for jet fuel is projected to grow close to 75% and increase availability of natural gas has similar long term shift away from heating oil. Refineries that can shift production to maximize diesel and jet fuel are best positioned to serve their markets. (Polasek & Mann, 2013) Compared to gasoline blending operations, blending of diesel, jet fuel, and other middle distillate products has been neglected since they are blended without component tanks. Many refineries in APAC and Europe have blend operations without component tanks and use excel with trial and error to obtain feasible solution. (Varvarezos) Usually, the refinery operations scheduling problem has been tackled by addressing the optimization of the three sub-operations, defined in previous chapter, independently of each other. This decentralized approach for scheduling optimization is not suitable for refineries where one or more component streams is directly blended without intermediate storage. Rundown blending optimization is challenging since many process units feed into the blender and fluctuations in the incoming flow rate and property has significant impact on finished product blend qualities and quantity. Additionally, rundown header is responsible for all the material that is transferred to it from the upstream process as seen in Figure 5.1. In a blending scenario with no intermediate tanks, the number of classic degrees of freedom is limited to the component rundown flow rate or the component qualities as manipulated variables. Thus if any changes happen in the blend recipe to meet finished product properties specifications, it would have a direct impact on the production unit operations. These interdependences of the blending operations and production unit operations require an integrated approach to scheduling. 107 Crude oil marine vessels Storage tanks charging tanks Crude Oil Unloading and Blending Production CDUs units Storage and other Production units Production Unit Operations Shipping points Blend headers Product tanks Product Blending and Delivery Figure 5.1 Graphic overview of a standard refinery system: includes no blend components storage tanks. It is imperative to model an integrated production unit operations and finished product blending-delivery problem to address the issue of rundown blending. The resulting comprehensive model is complex, hard to build and solve, and has received sparse attention in the literature. Moro et al. (1998) proposed a planning model for refinery diesel production where the emphasis is on blend relations. Pinto et al. (2000) proposed a planning and scheduling model for refinery operations. They presented the formulation based on discretization of time for production and distribution scheduling and their model included the features such as sequence dependent transition cost of products within oil pipeline.(C. Luo & Rong); Chunpeng Luo and Rong (2007) developed a two decisions levels scheduling model for overall refinery. The upper level optimization model is based on discrete time formulation and it is used to determine sequencing and timing of operations modes and to decide the quantities of materials produced/consumed at each operations mode. The upper level decisions level uses aggregated tanks storage capacity whereas the lower level uses heuristics to obtain a detailed schedule. They consider multi-purpose product tanks via an iterative procedure that allows to readjust aggregated tank capacity at lower 108 level by changing multi-purpose tank service mode and then to recalculate corresponding optimal solution at upper level. The logistics details are ensured through heuristics at the lower level. In previous chapter we have proposed a MILP model based on continuous-time representation for the simultaneous scheduling of production unit operations and end-product blending and delivery operations. The model incorporates quantity, quality, and logistics decisions related to real-life refinery operations. These logistics details involve minimum runlength requirements, fill-draw-delay, one-flow out of blender, sequence dependent switchovers, maximum heel quantity, and downgrading of the better quality product to the lower quality. Incorporating these logistics details in blend operations can yield significant improvement in productivity and efficiency. (Kelly, 2006) We proposed a set of valid inequalities that improves the computational performance of the model significantly in section 4.3.1. However, even with valid inequalities present, the scheduling model is still computationally prohibitive. It is imperative to develop global optimization strategy for an integrated production unit and finished product blending-delivery scheduling problem to address the issue of online blend units. The chapter is organized in the following way. Section 5.2 describes the integrated refinery operations problem. Section 5.3 presents solution strategy for the integrated problem, whereas section 5.4 defines relaxed sub-problems and presents methodology to construct novel heuristic approach to construct restricted Lagrangian sub-problems and feasible solution. Steps for obtaining tighter lower bounds are described in section 5.5 and detailed algorithm steps are presented in section 5.6. Section 5.7 provides results, and the chapter concludes with section 5.8. 5.2. Refinery Operations Scheduling The refinery operations scheduling MILP model proposed in Chapter 4 is used in this work to serve as a benchmark for the proposed decomposition algorithm. The formulation for the refinery operations scheduling problem is based on the continuous time representation and an idea of unit-specific event-points first introduced by M. G. Ierapetritou and C. A. Floudas (1998); 109 M.G. Ierapetritou and C.A. Floudas (1998); Ierapetritou et al. (1999). State-task network representation introduced by Kondili et al. (1993) is used in formulating the problem. Continuous time representation is preferable because it leads to lower number of binaries and constraints compared to discrete time representation and accurate start and finish timing of tasks. Discrete formulation requires the time horizon to be divided into smaller grids to acquire an acceptable approximation of timings which leads to increasing number of binaries. (Stefansson et al., 2011) In contrasts, continuous time formulation has more complicated structure where task can start and finish at any time point. (Floudas & Lin, 2004) The goal of the refinery operations scheduling is to maximize the production performance while minimizing the penalties subject to quantity, quality, and logistics giveaways and nonattainment. (Kelly, 2003a) The optimum values of the following variables in the system are driven by the shipment plan and accounts for the tradeoffs between costs of keeping inventory and changing run-modes: (i) sequencing of tasks for production units; (ii) each product pool that satisfies demand orders; (iii) duration of tasks at production units and duration of unloading tasks from product pools; (iv) inventory levels in component and product pools; (v) production rates for units and unloading rate for product pools; (vi) composition of material produced and consumed. 5.3. Solution Strategy The integrated refinery scheduling model (IP) presented in previous chapter gives rise to a large scale complex MILPs problem that requires specialized solution algorithms. We apply a Lagrangian decomposition (LD) algorithm to solve the integrated scheduling problem (IP) using an iterative procedure. The LD algorithm involves relaxing complicating constraints to the objective function by introducing Lagrange multipliers to form a relaxed version of a primal problem. In LD algorithm, we obtain lower bound and upper bound of the optimal value of (IP) at each iteration. 110 = start time of transfer to Blend Header P j ,n Tsn Tfn Pj, n Tsn B j ,n blend unit „j‟ in event point „n‟. Tfn Bj, n = end time of transfer to blend P s , j ,n Cfn unit „j‟ in event point „n‟. B s , j ,n Cfn = flow of component „s‟ to blend unit „j‟ in event point „n‟. Production Unit Scheduling Blending and Delivery Scheduling Figure 5.2 Spatial decomposition of a refinery operations network In this work, the integrated full-scale scheduling problem is decomposed into production unit scheduling problem (PSP) and blend scheduling problem (BSP) using spatial decomposition, as shown in Figure 5.2. Here, the network is split into two independent sub-problems; furthermore, each sub-problem can be further decomposed into smaller independent subproblems. The variables pertaining to the blend components flows amount, and start and finish times of these flows are also split alongside the connections (pipelines) in the network. Thus, two sets of linking variables, one set that belongs to the production unit scheduling sub-problem (PSP) and another set that belongs to the blend scheduling sub-problem (BSP), are present. To achieve this decomposition, coupling constraints are introduced to the integrated model (IP). Coupling constraints equate the split variables (flow amount, start and end time variables) of the blend components pipelines. Based on this spatial decomposition of the refinery structure, the main goal of the production unit scheduling sub-problem is to satisfy the demand requirement of final products that belong to set S B and demand of the blend components required by the subproblem. Similarly, the goal of the blend scheduling sub-problem is to satisfy the demand of the 111 finished blend products belonging to set S A by mixing the raw materials, supplied by (PSP), following the blending recipe and product property specifications. 5.4. Lagrangian Relaxation Framework and Sub-problems Lagrangian relaxation provides an efficient way for obtaining lower bounds for large scale MILPs. These problems are characterized by a set of complicating constraints whose removal to objective function using Lagrangian multipliers yields a relaxed problem that can be decomposed into smaller independent sub-problems that are easier to solve. In classical LD, lower bound is obtained from the solution of the relaxed problem and upper bound is obtained by constructing a feasible solution based on a solution to the relaxed problem. Then the multipliers are updated along a sub-gradient direction. The general framework of the classical Lagrangian decomposition (LD) algorithm is given in Figure 5.3. Here m is used to account for the algorithm iterations, Z L is lower bound, m and Z U is upper bound. u Tsj,t m iteration. (m) , u Tfj,t (m) , and usCf, j ,t (m) are the Lagrange multipliers at the mth 112 Initialize Lagrange multipliers , , , Solve Relaxed Problem Obtain lower bound ( ) Construct feasible solution Obtain Upper bound ( ) No Update Lagrange multipliers , , , Yes Done Figure 5.3 Framework of classical Lagrangian Decomposition Algorithm In our work we develop a modified LD algorithm to suit the application to refinery continuous production processes, especially for refinery that have a blend header receiving at least one blend component stream directly from upstream processes without component storage. We formulate restricted relaxed sub-problems using the solution of Lagrangian relaxed sub-problems to obtain tighter lower and upper bounds. The details of each step in Lagrangian algorithm are given in the following subsections. 5.4.1. Relaxed Problem The connections between production units and blend units are cut off by splitting the pipelines between them since the refinery network does not contain blend component tanks. Thus, the variables JJf s , j , j , n , s Sbc , j J sp J PU , j J sc J BU , n N pertaining to the blend component flow are replaced with CfnsP, j , n , CfnsB, j , n , s Sbc , j J BU J sc , n N . Here, J BU is a set of blend units and J PU is a set of units present in (PSP). Furthermore, the time sequence 113 constraints which enforce that the start and finish time of the production units which produce blend component ( s ) should be same as the blend unit which consumes the blend component ( s ) are eliminated from the relaxed problem. In place of these sequence constraints, we introduce variables Tsn j , n , Tfn j , n , j J BU , n N for start and finish time of flow from production unit scheduling problem to a blend unit b . The variables pertaining to the blend component flow and start/end times of flow for all split connections between production unit operations and blend units are duplicated. The model has two sets of linking variables, one set that belongs to the production unit scheduling sub-problem Cfn , Tsn Pj , n , Tfn Pj , n Cfn , Tsn Bj , n , Tfn Bj , n . All other variables in model (IP) are called non-linking variables since they P s, j ,n B s, j ,n and another set that belongs to the blend scheduling sub-problem are separate for both sub-problems. Coupling constraints (1)-(3) are introduced to equate the split variables (flow amount, start and end time variables) for every event-point n . TsnPj , n TsnBj ,n , j J BU , n N (1) TfnPj , n TfnBj ,n , j J BU , n N (2) CfnsP, j , n CfnsB, j , n , j J BU , s S cj , n N (3) To obtain the Lagrangian relaxation of the original problem, the complicating constraints (1-3) are relaxed using Lagrange multipliers ( u ) and considered in the objective function as shown in equation (4). Ln u z B P u Tsn j , n Tsn j , n Tsn j , n j J BU ,n B P uTfn j , n Tfn j , n Tfn j , n j J BU ,n jJ BU , sS cj , n B P usCfn , j , n Cfns , j , n Cfns , j , n (4) The objective function (4) of the relaxed problem is decomposable into smaller sub-problems corresponding to the production unit operations and the finished product blending and delivery operations, which are relatively easier to solve. The model (P1-n) includes objective function ( LBn ) and all the constraints and variables pertaining to the finished product blending and delivery 114 operations, whereas, the model (P2-n) includes objective function ( LPn ) and all the equations and variables pertaining to the production unit operations. B Tfn B B minimize LBn u z uTsn usCfn j , n Tsn j , n u j , nTfn j , n , j , n Cfns , j , n jJ BU , n jJ BU , n jJ BU , sS cj , n s.t. constraints corresponding to finished product blending and delivery operations (P1-n) P Tfn P P minimize LPn u z uTsn usCfn j , n Tsn j , n u j , nTfn j , n , j , n Cfns , j , n jJ BU , n jJ BU , n jJ BU , sS cj , n s.t. constraints corresponding to production units scheduling operations (P2-n) The models (P1-n) and (P2-n) are independent and can be solved in parallel. Their solutions will provide a feasible solution to the original problem when the coupling constraints (1-3) are satisfied for all event-points. Figure 5.4 shows a typical solution obtained by solving (P1-n) and (P2-n) for problem that has one blend unit, say b1. Gantt chart shows the flow amount and start/end times of one of the blend component ( s1 ) for blend unit b1. As you can see, blend components are being supplied from production units to the blend header and consumed by the blend header at the same time. However, if one were to look at Gantt chart closely, the eventpoints for the component streams are not the same. The complicating constraints (1)-(3) are not satisfied for event-points n 2 and n 3 . n1 n2 n4 P1-n n1 n3 n4 P2-n Time 115 Figure 5.4 Gantt chart for a blender receiving component s. This kind of situation arises frequently when comparing solutions of (P1-n) and (P2-n) across event points due to the nature of unit specific event points and continuous time representation used in the model. To meditate this kind effect, we introduce a time-point t that has one-to-one correspondence with an active event-point n and cardinality of set T is equal to the cardinality of set N . To eliminate the occurrence of non-active time-points before active time-points, additional constraints are included to enforce that active time-points always occur before non-active time-points. To compare material flow between production operations and blend operations at each time points, a new set of variables Cf s , j ,t , Ts j ,t , Tf j ,t are introduced. The coupling constraints presented in equation (1)-(3) are replaced by (1b)-(3b). Ts Pj ,t Ts Bj ,t , j J BU , t T (1b) Tf jP,t Tf jB,t , j J BU , t T (2b) Cf sP, j ,t Cf sB, j ,t , j J BU , s S cj , t T (3b) The Lagrangian relaxation objective function Ln (u) given in equation (4) is updated to L(u) given in equation (4b), L u z jJ BU ,t u Tsj ,t Ts Bj ,t Ts Pj ,t jJ BU ,t u Tfj ,t Tf jB,t Tf jP,t jJ BU , sS cj ,t usCf, j ,t Cf sB, j ,t Cf sP, j ,t (4b) Where the variables Cfns , j , n , Tsn j , n , Tfn j , n are replaced with variables Cf s , j ,t , Ts j ,t , Tf j ,t . The relaxed sub-problems (P1-t) and (P2-t) are obtained by decomposing L(u) . minimize LB u z uTsj ,t Ts Bj ,t uTfj ,t Tf jB,t usCf, j ,t Cf sB, j ,t jJ BU , t jJ BU ,t jJ BU , sS cj ,t s.t. constraints corresponding to finished product blending and delivery operations (P1-t) 116 minimize LP u z u Tsj ,t Ts Pj ,t u Tfj ,t Tf jP,t usCf, j ,t Cf sP, j ,t jJ BU , t jJ BU ,t jJ BU , sS cj ,t s.t. constraints corresponding to production units scheduling operations 5.4.1.1. (P2-t) Relaxed Blend Scheduling Problem Relaxed blend scheduling problem includes all the constraints and variables of the relaxed sub-problem (P1-t) and in addition to those constraints and variables; we introduce a new continuous 0-1 variable x Bj, n ,t to connect the time-points t and event-points n . x Bj , n ,t wvi , j , n , j J BU , n N , t T , n t iI j x Bj , n,t 1 x Bj , n,t 1 x Bj , n ,t , j J BU , n N , t T , n t (6) x Bj , n,t , j J BU , n N , t T , n t (7) t t ,t n n n , n t x Bj , n ,t wvi , j , n iI j x n t B j , n ,t (5) t t ,t n x Bj , n,t n n , n t x Bj , n,t , j J BU , n N , t T , n t x Bj , n ,t , j J BU , t T , t T , t t nt (8) (9) x Bj , n,t x Bj , n,t 1, j J BU , n N , n N , n n, t T , t T , t t , n t , n t , n t , n t (10) x Bj , n,t 0, j J BU , n N , t T , n t (11) Constraint (5) states that variable x Bj, n ,t is non-zero at time-point t only if a blend unit j is active at event-point n. Here wvi. j.n is a binary assignment variable, 1 if task i is active in unit j at eventpoint n, and 0 otherwise. Constraints (6) and (7) enforce that each blend unit j has one-to-one correspondence between a time-point and an event-point. Equation (8) states that if the blend unit is active at event-point n then it must correspond to some time-point t. Constraint (9) restrict that time-point t t should be assign an event-point before time-point t. Similar to constraint (10), 117 n n should be assigned to a time-point t t before assigning n to t. Since, we want to restrict inactive time-points ( x Bj, n,t 0 ) to happen after active time-points ( x Bj,n,t 1 ), we add equation (11). The decoupled flow and time variables at each time-point are defined as follow: Ts Bj ,t Tsi , j , n UH 2 wvi , j , n x Bj , n ,t , j J BU , i I j , n N , t T , n t (12a) Ts Bj ,t Tsi , j , n UH 2 wvi , j , n x Bj , n ,t , j J BU , i I j , n N , t T , n t (12b) Tf jB,t Tf i , j , n UH 2 wvi , j , n x Bj , n ,t , j J BU , i I j , n N , t T , n t (12c) Tf jB,t Tf i , j , n UH 2 wvi , j , n x Bj , n ,t , j J BU , i I j , n N , t T , n t (12d) Cf sB, j ,t Uif s , j , n M 1 x Bj , n ,t , j J BU , s S cj , n N , t T , n t (13a) Cf sB, j ,t Uif s , j , n M 1 x Bj , n ,t , j J BU , s S cj , n N , t T , n t (13b) Ts Bj ,t UH x Bj , n ,t , j J BU , t T (14a) Tf jB,t UH x Bj , n ,t , j J BU , t T (14b) n t n t Cf sB, j ,t M x Bj , n ,t , j J BU , s S cj , t T n t (15) Constraints (12a)-(12d) assign start and end times to time-point t using the corresponding eventpoint n. Similarly, blend components flow amount is assigned to time-point t using equation (13a) and (13b), where M is a big M term. If a time-point is not active then the variables pertaining to blend components flows are equal to zero as defined by equations (14a), (14b) and (15). If refinery under study contains multiple blend units, then the relaxed blend scheduling problem may be decomposed into multiple independent blend scheduling problems. Each of these independent blend scheduling problems must blend different components, more specifically, any two independent scheduling problems must not share a common blend component. x n t B j , n ,t 1, j J BU , t T (16) 118 x t n B j , n ,t wvi , j , n , j J BU , n N (17) iI j To improve the computational performance of the decomposed problems, valid inequalities (16) and (17) are added to the blend scheduling models. Inequality (16) states that for a given timepoint t, continuous 0-1 variable x Bj, n ,t can only correspond to one event-point and equality (17) states that if blend unit is active at given event-point n then there must be a time-point t corresponding to n. 5.4.1.2. Relaxed Production Unit Scheduling Problem Continuous 0-1 variables x Pj, n ,t and a j ,n are introduced to define split variables belonging to production unit scheduling problem in terms of time-point t instead of event-point n. To this goal, we first introduce equations (18a) and (18b). a j ,n j J iI j , S cj Sip PU wvi , j , n , j J BU , n N , iI j , S cj Sip wvi , j , n a j , n , j J BU , j J PU , j J seq j , n N , (18a) (18b) Binary variable a j ,n takes the value of 1 if at least one production task produces a blend component at event-point n that can be consumed by blend unit j as stated by constraint (18a). Equation (18b) restricts a j ,n to 0 if production operations are not producing blend components at event- point n for blend unit j. Similarly to blend scheduling problem, production unit scheduling problem has equations (19)(25) to define 0-1 continuous variable x Pj, n ,t . x Pj , n,t a j ,n , j J BU , n N , t T , n t x Pj , n,t 1 x Pj , n,t 1 (19) x Pj , n ,t , j J BU , n N , t T , n t (20) x Pj , n,t , j J BU , n N , t T , n t (21) t t ,t n n n , n t 119 x Pj , n ,t a j , n x n t P j , n ,t t t ,t n x Pj , n ,t n n , n t x Pj , n ,t , j J BU , n N , t T , n t x Pj , n ,t , j J BU , t T , t T , t t nt (22) (23) x Pj , n,t x Pj , n,t 1, j J BU , n N , n N , n n, t T , t T , t t , n t , n t , n t , n t (24) x Pj , n,t 0, j J BU , n N , t T , n t (25) Constraint (19) states that variable x Pj, n ,t is non-zero at time-point t only if there is at least one production unit operational at event-point n that can produce a blend component necessary for blend unit j. Constraints (20) and (21) enforce that there is one to one relationship between a time-point and an event-point for a every connections between production unit scheduling problem and blend unit j. Equation (22) ensures that if there is a blend component production happening event-point n then it must correspond to some time-point t. Constraint (23)-(25) are same as constraints (9)-(11) present in the blend scheduling problem. wvi , j , n a j , n , j J BU , s S cj , n N j J PU J sp , iI j I sp (26) Since there are no blend component tanks present, blend unit must receive all blend components from production scheduling problem at the same time for inline continuous blending. Constraint (26) requires that if at least one blend component is produced at event-point n for a blend unit j then all other blend components for the blend unit must be produced at event-point n. This requirement is valid because in a typical refinery, a blend unit produces multiple products by mixing a fixed set of blend components using different blend recipes. That is, j J BU , I I j , the blend components that can be consumed by different tasks at blend unit j gives us S cj Sic . The goals of the production unit scheduling problem is to produce hydrocarbon products ( S B ) that are not stored in tanks but are supplied straight to the market and to produce blend components that are needed by blend units to produce finished blend products. Since there are no tanks available to store blend components, these products (blend components) are supplied 120 straight to the market (here market is blend unit). Thus, the product set ( S B ) includes the blend components and other hydrocarbon products that will be used as feedstock in chemical industry. Constraints relating the split variables (flow, start and end times) of the production scheduling problem to time-points are given below. Ts Pj ,t UH x Pj , n ,t , j J BU , t T (27a) Tf jP,t UH x Pj , n ,t , j J BU , t T (27b) n t n t Ts Pj ,t Tsi , j , n UH 2 wvi , j , n x Pj , n,t , p c j J PU , j J BU , j J seq j , i I j , n N , t T , n t , Si S j Ts Pj ,t Tsi , j , n UH 2 wvi , j , n x Pj , n,t , p c j J PU , j J BU , j J seq j , i I j , n N , t T , n t , Si S j Tf jP,t Tfi , j , n UH 2 wvi , j , n x Pj , n,t , p c j J PU , j J BU , j J seq j , i I j , n N , t T , n t , Si S j Tf jP,t Tfi , j , n UH 2 wvi , j , n x Pj , n,t , p c j J PU , j J BU , j J seq j , i I j , n N , t T , n t , Si S j (28a) (28b) (28c) (28d) Constraints (27a) and (27b) states that start and end times are zero if material is not flowing to a blend unit j’ at time-point t. If production unit operations are producing blend components at time-point t, then equations (28a)-(28d) are used to assign start and end times to blend component flow to a blend unit j’. All production and blend units in the refinery are continuous processes and since there is no storage between the production unit scheduling operations and the blend units, material produced by the production unit scheduling problem is directly consumed. Thus production capacity of blend units and blend recipes of each blend component should also be taken into consideration when the components are supplied by production unit operations. Parameters RBmax and RBmin are j j maximum and minimum blend rate of a blend unit j, respectively and are calculated using 121 equations (29a) and (29b). The maximum and minimum blend recipe of component s in blend unit j, bsmax and bsmin , j respectively, are determined from Equations (30a) and (30b). ,j RB max max Rimax j J BU j , j , (29a) RB min min Rimin j J BU j , j , (29b) max bsmax j J BU , j max s , i , (30a) min bsmin j J BU , j min s , i , (30b) iI j iI j iI j iI j The blend component production is set to zero by equation (31a) if there is no production happening at time-point t for blend component s. Constraints (31b) and (31c) determines the individual component flow amount from the production unit scheduling problem to a blend unit j at time-point t. max P Cf sP, j ,t bsmax j J BU , s S cj , t T , j RB j UH x j , n , t , nt s , j , n max P bsmax j J BU , s S cj , n N , t T , n t , j RB j UH 1 x j , n , t , (31b) s , j , n max P bsmax j J BU , s S cj , n N , t T , n t , j RB j UH 1 x j , n , t , (31c) Uof Cf sP, j ,t j J sp Cf sP, j ,t Uof j jsp (31a) Cf P s , j ,t P RB max j J BU , t T Tf jP,t Ts Pj ,t RB max j j UH 1 x j , n ,t , n t (32a) Cf P s , j ,t P RB min j J BU , t T Tf jP,t Ts Pj ,t RB min j j UH 1 x j , n ,t , n t (32b) sS cj sS cj max P Cf sP, j ,t bsmax j J BU , s S cj , t T Tf jP,t Ts Pj ,t bsmax, j RB max , j RB j j UH 1 x j , n ,t , nt (32c) min P Cf sP, j ,t bsmin j J BU , s S cj , t T Tf jP,t Ts Pj ,t bsmin, j RB min , j RB j j UH 1 x j , n ,t , n t (32d) 122 Constraints (32a) and (32b) enforce that total amount of material supplied to a blend unit j at time-point t is bounded by maximum and minimum production rate of corresponding blend unit and production of each blend component is bounded by equations (32c) and (32d). Similar to valid inequalities (16) and (17) included in blend scheduling sub-problem, we introduce constraints (33) and (34) to production unit scheduling problem. Inequality (33) states that for a given time-point t, continuous 0-1 variable x Pj, n ,t can only correspond to one event-point. If at event-point n production unit operations are producing materials for blend unit j, then there must be at least one time-point t corresponding to n as stated by equality (34). x P x P j , n ,t n t t n j , n ,t 1, j J BU , t T a j , n , j J BU , n N (33) (34) 5.4.2. Construction of a Feasible Schedule Traditionally an upper bound for the Lagrangian decomposition is obtained by fixing certain binary variables of the original full-scale integrated MILP problem (IP) based upon the solution of the relaxed sub-problems. However the solution derived by solving relaxed sub-problems ((P1t) and (P2-t)) are usually infeasible in each iteration because the complicating constraints are are violated. Therefore, a heuristics procedure or some other methods for generating feasible solution is necessary. In this work, feasible solution at each iteration is obtained by fixing some of the binary variables to 1 and to decide which binaries to fix, we propose using solution of the relaxed sub-problems to build new restricted relaxed sub-problems. These restricted relaxed subproblems‟ solution will always provide feasible fullspace solution unlike relaxed sub-problems. In proposed decomposition strategy, complicated variables are expressed in terms of eventpoint t and not event-point n , thus the schedule obtained solving relaxed problem (P1-t) and (P2t) would lead to an infeasible solution of the (IP). In most instances the infeasibility in full-scale model occurs because the fundamental operational rule of the refinery is violated because material balance constraints are not satisfied for blend units or production units that produce 123 blend components. The refinery system under study doesn‟t contain blend component tanks thus blend operations and corresponding production unit operations must operate concurrently to satisfy mass balance constraints of rundown streams. Relaxed sub-problems (P1-t) and (P2-t) defined in previous section are solved independent of each other and their solutions are compared at time-point t . Thus at some iteration m , the production operations scheduling model (P2-t) may produce a solution where blend components are supplied to a blend unit j during eventpoints n 2 and n 4 , that is x Pj,2,1 1 and x Pj,4,2 1 , while relaxed blend scheduling model (P1-t) solution has the blend unit j is receiving components during event-points n 1 and n 3 , that is x Bj,1,1 1 and x Bj,3,2 1 . Hence at if the binary variables wv i, j, n i, j ,n in the original integrated model (IP) are fixed based upon the solution of relaxed sub-problems, it would lead to an infeasible solution. Infeasibility would occur because in the refinery, for two sequential units without storage any between them must operate at the same event-point. Hence sub-problems must provide a solution where a blend unit j J BU is receiving and blending components at the same event-points. That is, both x Pj, n ,t and x Bj, n ,t have to be nonzero at event point n to provide feasible solution of full space problem. Before the steps to build restricted relaxed sub-problems are presented, we first state assumptions behind proposed method to derive restricted relaxed sub-problems. We assume that no parallel blend units are present in the refinery and no two blend units share a particular blend component. These assumptions are valid because, in this work, the refinery does not have any blend component tanks present. The mathematical representation of these assumptions is given below as: S cj S cj , K jpk K jpk j J BU , j J BU , j j 124 Thus, blend scheduling problems (BSP) can be decomposed into sub-problems containing only one blend unit. Here each (BSP) sub-problem defined as BSP jJ BC has its own set of blend components, materials, and tanks corresponding a blend unit j J BC . Now we analyze the structure of the (PSP), particular for a refinery that has more than one blend unit and the (BSP) can be decomposed into independent sub-problems. We check if the production units that supply components to different (BSP) sub-problems are connected or not. Production units are said to be “connected” if they all have to be operational at the same time which corresponds to being active at same event-points so that the continuous process operational rule is not violated for the (PSP). We assume in our work that all the production units which produce blend components for different blend units are connected. Based upon this assumption, all independent sub-problems of the (BSP) must also be active at the same event-point to obtain a feasible solution to the integrated model (IP). A two steps approach is developed to determine an upper bound for the algorithm which is obtained from a feasible solution to the original full-scale model. In step 1 a restricted relaxed problem is solved to obtain a solution that has production units active at the same event-points as the blend units. Using the solution of step 1, in step 2 a feasible solution to (IP) is realized. 5.4.2.1. Restricted Relaxed Problem Upon solving relaxed sub-problems presented in section 5.4.1, we obtain schedule for blending operations and production unit operations, respectively. To build a restricted relaxed problem using the optimal solution of relaxed problem, the following steps are proposed: Step 1: From the optimal relaxed sub-problems solution, determine over the scheduling horizon which tasks, units, and storage tanks are active for at least one event-point. Step 2: Pick a primary sub-problem from PSP BSP j , j J BU by analyzing solution of the corresponding relaxed sub-problems. Determine active event-points during which the primary sub-problem is either supplying or receiving blend components. 125 Step 3: Obtain restricted relaxed problem by fixing some of the binary variables in corresponding relaxed problem to 1 at active event-points. Step 1: Determine parameters from an optimal solution of relaxed sub-problems. Using equations (35a)-(35d), we can calculate active task for units ( wvia, j ), the flow in and out of tanks ( insa, j ,k and outsa, k , j ), and the changeovers from task i to i at unit j , respectively over the scheduling horizon. Parameter ia,i, j is calculated using equation (35e) which gives the total number of time a task i is change to a task i at unit j . a i, j wv a s, j ,k in out 1 if 0 if wv wv 1 if 0 if in in a s,k , j a i ,i , j 0 i, j ,n 0 j J , i I j (35a) n s , j , k .n 0 s, j ,k ,n 0 n s S , k K s , j J kpk (35b) n out out 1 if 0 if 1 if 0 if i, j ,n n s,k , j ,n 0 s,k , j ,n 0 n s S , k K s , j J kkp (35c) n i ,i , j , n 0 i ,i , j , n 0 n j J , i I j , i I j (35d) n ia,i , j i ,i , j , n j J , i I j , i I j n (35e) Step 2: From the solution of relaxed sub-problems, the total number of active time-points for each Lagrangian decomposed problem are calculated as: Tat j x t ,n t j , n ,t , j J BU Since we assume that the production units which produce blend components for different blend units are connected, for (PSP), Tat P =Tat Pj Tat Pj , j J BU , j J BU , j j 126 The total numbers of active-points for each sub-problem BSP j containing blend unit j are given as Tat Bj . The maximum number of active time-points is determined by finding maximum among (BSP) and (PSP) as: Mat max Tat P Tat Bj , j J BU If there is more than one sub-problem that has the total active time-points equal to Mat , then the following rules are used to determine the primary problem: Rule 1: If Tat Bj Mat and sub-problem BSP j includes multipurpose blend unit j then the primary problem is ( PP BSP j ). Rule 2: If Tat Bj Mat and sub-problem BSP j doesn‟t include multipurpose blend unit j then the blend scheduling sub-problem is chosen as the primary problem ( PP BSP j ). Rule 3: If Tat Bj Mat and Tat P Mat then (PSP) is selected as the primary problem ( PP PSP ). When determining the primary problem, priority is given to a problem that satisfy rule 1; however if rule 1 is not satisfy then we check rule 2 and subsequently rule 3. The solution of primary problem is used to determine parameter Aen which represents active event-point. 1 if wvi , j , n 0 iI j Aen , 0 if wvi , j , n 0 iI j j J BC , BSP j PP 1 if c p wvi, j,n 0 j J PU , jJ seq j , iI j , S j Si Aen , 0 if PU seq c p wvi , j , n 0 j J , jJ j ,iI j , S j Si j J BC , PSP PP (36a) (36b) Parameter Aen is determined by equation (36a) if the primary problem is a sub-problem BSP j otherwise it is determined by equation (36b). 127 Step 3: Using parameters wvia, j , insa, j ,k , outsa,k , j , ia,i, j , ia,i, j and Aen we restrict the solution space of the relaxed sub-problems so that a solution of the restricted problem have production operations producing blend components at the same event-point as the blend units are mixing.. We fix binary variables { wvi , j ,n , ins , j ,k ,n , outs , k , j ,n , i ,i, j ,n } as shown equations (37a)-(37h). wvi , j , n wvia, j Aen , j J , j J m , i I j , n N (37a) wvi , j , n 0, j J m , i I j , n N , Aen 0 wvia, j 0 (37b) wvi , j , n Aen , j J m , n N , Aen 0 (37c) iI j , wvia, j 0 n , Aen 0 wvi , j , n wvia, j , j J m , i I j , wvia, j 0, ia,i , j 0 i ,i i ,i , j , n 0, j J , i I j , i I j , n N , Aen 0 ia,i , j 0 n , Aen 0 i ,i , j , n ia,i , j , j J , i I j , i I j , i i , n N , ia,i , j 0 (37d) (37e) (37f) ins , j , k , n insa, j , k Aen , k K , k K m , k K p , j J kpk , j J m , i I j , s Sk , n N (37g) outs , k , k , n outsa, k , j Aen , k K , k K m , k K p , j J kkp , j J m , i I j , s Sk , n N (37h) Equation (37a) fixes binary assignment variable for all units except for multipurpose units. For multipurpose units, equations (37b)-(37f) are used, where equation (37b) fixes assignment variable wvi , j ,n to zero if the task i has not been active in relaxed problem or if the event-point n is not active in primary problem. Constraint (37c) forces unit to be active when Aen 1 and constraint (37d) enforces that task i needs be active in unit j if wvia, j 1 and no tasks changeovers are happening i , i a i , i , j 0 over the time horizon. Equation (37e) fixes continuous 0-1 variable i ,i, j ,n to be zero at event-point n if Aen 0 and the total number times the changeover from task i to i can happen in unit j is fixed to ia,i, j in constraint (37f). Binary variables ins , j ,k ,n and outs , k , j ,n are fixed in restricted relaxed problem as shown in equations (37g)(37h) for all units and tanks that do not perform multipurpose tasks and store final product tanks. 128 Since the blend products demand is to be satisfied before its due date, we do not fix any variables associated with the product tanks as not to restrict solution space of the restricted relaxed problem too much that product demands do not get satisfied. This situation can occur because in refinery under study, the products tanks can‟t load and unload products simultaneously. Thus, if we fix the binary variable associated with loading and unloading of product tanks based upon the solution of the relaxed problem, it may cause infeasibility in the restricted relaxed problem. Restricted relaxed problem is built and solved for every sub-problem except for the subproblem that is chosen as the primary problem. Solution of the restricted relaxed problem gives us consistent solution across (BSP) and (PSP) in terms of event-points so that the continuous process material balance is not violated. 5.4.2.2. Upper Bounding Problem We fix many of the binary variables { wvi , j ,n , ins , j ,k ,n , outs , k , j ,n , lk ,o,n } in problem (IP) to the values of the corresponding binary variables obtained from the solution of restricted relaxed subproblems and obtain a MILP model (IP-U). The solution of the restricted relaxed problem is { wvir, j , n , insr, j , k , n , outsr, k , j , n , lkr,o, n } and based on non-zero values of these binaries, corresponding variables in (IP) are fixed. Only certain binary variables in problem (IP) are fixed to 1 and no binaries are fixed to 0 so that a better upper bound can be attained in earliest iterations in the LD algorithm. Equations (38a)-(38d) show how the variables are fixed in the problem (IP) to obtain a MILP model (IP-U). wi , j , n 1, j J , i I j , n N , wvir, j ,n 0 (38a) ins , j , k , n 1, k K , j J kpk , i I j , s Sk , n N , insr, j ,k 0 (38b) outs , k , k , n 1, k K , j J kkp , i I j , s Sk , n N , outsr,k , j 0 (38c) lk ,o, n 1, k K p , j J kpk , i I j , s Sk , n N , lkr,o,n 0 (38d) Solving model (IP-U) yields an upper bound on the solution of (IP). The model (IP-U) is always feasible because the model includes demand, due-dates, and blend logistics giveaways. 129 5.5. Strengthening the Lagrangian Relaxation There are several constraints, redundant or unnecessary in (IP) that can be added to the relaxed problem and the restricted relaxed problem to strengthen the performance of the Lagrangian algorithm. To improve the lower bound, we propose a pre-processing step before the Lagrangian algorithm and add constraints related to product demands to the decomposed problems in the algorithm. A common feature of the preprocessing steps and constraints strengthening lower bound is that they in some sense recreate what has been relaxed, namely the network structure of the refinery, but without destroying the separability of the Lagrangian sub-problems. Furthermore, using the solution of the best upper bound, in the decomposed problems, penalty variables relating violations of demands, due dates, and blend logistics can be fixed to zero. 5.5.1. Preprocessing Step Because the refinery is decomposed into (BSP) and (PSP), production capacity limitation for the blend components in (PSP) is not taken into consideration while solving (BSP). Therefore, we present a preprocessing step that determines the maximum production rate for each blend component s and maximum production capacity for a set of blend components that can be mixed by a blend unit. minimize zsPP 100 Uof s , j , n , s Sbc PU p jJ , jJ s , nN (39) minimize z PP 100 Uof s , j , n , j J BU j j J PU , sS cj S jp , nN (40) The preprocessing problem (PC-s) and (PC-j) are based upon the relaxed sub-problem corresponding to production unit operations (PSP). The model (PC-s) is obtained by replacing objective function of relaxed problem with that given in equation (71) to determine maximum production rate of each blend component. Another preprocessing problem (PC-j) is presented to calculate the maximum rate at which production units can supply materials to each blend unit. 130 The objective function for (PC-j) is presented in equation (72). Problem (PC-s) is solved for each blend component and problem (PC-j) for each blend unit before the Lagrangian algorithm begins. Furthermore these problems are independent of each other and can be solved in parallel. Upon solving the pre-processing problems, we obtain values of variables Cf sP, j ,t , Ts Pj ,t , Tf jP,t , x Pj , n ,t s, j, n, t from the optimal solution. RP Cf sP, j ,t x Pj , n,t BU c n max , j J , s S j P P t Tf Ts j ,t j ,t (41a) RPjmax Cf sP, j ,t x Pj , n,t sS c n BU max j P , j J P t Tf j ,t Ts j ,t (41b) max s The maximum rate of production of component s is determined by equation (41a) using the optimal solution obtained upon solving corresponding preprocessing problem (PC-s). Similarly, the maximum rate for set of blend components supplied to blend unit j is determined by equation (41b) using the optimal solution values of corresponding preprocessing problem (PC-j). 5.5.2. Strengthening Lower Bound To strengthen the Lagrangian decomposition algorithm, the information obtained in preprocessing step is applied to update the blend component production rate upper bounds in equations (13a), (13b), and (15) for relaxed and restricted relaxed blend scheduling problems. Computationally inefficient big M term is replaced with tighter bound upper RPsmax as shown in equations (42a)-(42c). Cf sB, j ,t Uif s , j , n RPsmax 1 x Bj , n ,t , j J BU , s S cj , n N , t T , n t (42a) Cf sB, j ,t Uif s , j , n RPsmax 1 x Bj , n ,t , j J BU , s S cj , n N , t T , n t (42b) Cf sB, j ,t RPsmax x Bj , n ,t , j J BU , s S cj , t T (42c) n t 131 Cf sB, j ,t RPsmax Tf jB,t Ts Bj ,t RPsmaxUH 1 x Bj , n,t , j J BU , s S cj , t T n t Cf B s , j ,t sS cj RPjmax Tf jB,t Ts Bj ,t RPjmaxUH 1 x Bj , n ,t , j J BU , t T n t (42d) (42e) The amount of blend component processed by blend unit j at time-point t is bounded by the maximum production rate of the component and blend unit‟s process time as enforced by constraint (42d). Similarly, constraint (42e) states that the total bending rate is bounded by the maximum rate at which material is supplied to a blend unit j. iI sp , wvi , j , n 1, s S A S B , Amt min 0 (43) j J i , n iI j , i I j ,i i , n i ,i , j , n 1 1, j J M J BU sS jp , iI sp , iI j Here parameter Amt min sS A , oOs Do, s sS A , k K s (44) stos , k . Constraint (43) enforces that tasks that produce product s must be active at least once over the scheduling time horizon and constraint (44) enforces minimum number of changeover of tasks at multipurpose blend units. Constraint (43) is added to both (PSP) and (BSP); however, equation (44) is only present in (BSP). Uof s , j , n Amt min pen sSbc , jJ sp , n N minimize LPm u z C11 pen u Tsj ,t Ts Pj ,t uTfj ,t Tf jP,t usCf, j ,t Cf sP, j ,t jJ BU ,t jJ BU ,t jJ BU , sS cj ,t (45) (46) Constraint (45) enforces that (PSP) must produce at least minimum amount of blend components required for (BSP) so that demand of blend products can be satisfied. If due to production capacity limitation, (PSP) cannot produce enough blend components, then we add positive penalty variable pen in equation (45) which is penalized in the objective function as shown in equation (46). The objective function of relaxed and restricted relaxed problems is now replaced by the one given in equation (46). 132 5.5.3. Construction of lower bound from the best upper bound Due to production capacity limitation, the refinery might not be able to satisfy the demand of the products on time, might not respect all the blend logistics requirements, and may not be able to meet product quality specifications. To obtain a feasible schedule that can be implemented in the refinery, we have included, demand, due dates, and blend logistics giveaways in all scheduling models. These violations are heavily penalized in the objective function. Equation (47a) gives the demand and due violations cost and equation (47b) gives total violation cost by including the product downgrading and minimum heel violation penalty to the demand violation cost. We also obtain the cost of task changeovers for multipurpose blend units from feasible solution as given in equation (47c). l 12 u violatndem c11 o dg o co dg o oO oO violatntotal violatndem c sS B 13 s k K , sS k , s S k , s s , n m chgovr rg s co14Tearlyo c15 o Tlateo oO ck9K m std s , s , k , n oO c10 s , k mhs , k , n k K , sS k , n m ci6,i i ,i , j , n jJ m , iI j , i I j , i i , n (47a) (47b) (47c) For a minimization problem, Z L Z opt Z U thus, if the best upper bound obtained at the current iteration provides a solution without any violations, then the optimal schedule of the original integrated problem will not have any violations. Furthermore, in refinery, an optimal solution with lowest task changeovers cost is desired if there are no demand and due-date violations. The desired changeovers at multipurpose blend units are from task producing high quality blend product to tasks producing low quality products so that the products produced during changeover U transition are not discarded. Thus, we can conclude that chgovr opt chgovrU , if violatntotal 0. dg , dg l o u o U , rg s , Tearlyo , Tlateo 0, if violatndem 0 (48a) dg , dg u o U , rg s , Tearlyo , Tlateo , std s , s , k , n , mhs , k , n 0, if violatntotal 0 (48b) l o 133 U ci6,i i ,i , j , n chgovr U , if violatntotal 0 jJ , iI j ,i I j ,i i , n m (48c) U We can produce tighter lower bound by considering value of parameter violatntotal obtained from the best upper bound solution for a given set of Lagrangian multipliers. If best upper bound obtained satisfies finished products demand and their due-dates, violatnUdem 0 , then we fixed variables associated with demands and due-dates to zero as given by equation (48a). Similarly, in constraint (48b), variables associated with demand and blend logistics violations are U fixed to zero if the best upper bound solution where violatntotal 0 . Furthermore, total changeover cost for multipurpose blend units is bounded by the total changeover cost ( chgovrU ) using equation (48c). 5.6. Lagrangian Decomposition Algorithm In this section, proposed Lagrangian decomposition algorithm is presented in detail. This algorithm differs from classical Lagrangian decomposition in the way that the lower bound and Lagrange multipliers are obtained. In classical Lagrangian decomposition (classical-LD), the lower bound is always obtained from the Lagrangian relaxed problem where as in the proposed restricted Lagrangian decomposition (restricted-LD) the tighter lower bound is chosen among the relaxed problem and the restricted relaxed problem. Before presenting the details of the algorithm, the scheduling models used in the algorithm are defined first. The relaxed blend scheduling model (P1) consists of tasks, units, materials, tanks relating to (BSP) and is defined as below: minimize LB u z u Tsj ,t Ts Bj ,t u Tfj ,t Tf jB,t usCf, j ,t Cf sB, j ,t jJ BU , t jJ BU , t jJ BU , sS cj , t s.t. constraints (5-11),(12a-12d),(14a-14b),(16-17),(42a-42e),(43-44),(48a-48c), and all the constraints corresponding to finished product blending and delivery operations in model (IP) (P1) 134 The relaxed production unit scheduling model (P2) consists of tasks, units, materials, tanks relating to (PSP) and is defined as below: minimize LPm u z C11 pen u Tsj ,t Ts Pj ,t u Tfj ,t Tf jP,t usCf, j ,t Cf sP, j ,t jJ BU , t jJ BU , t jJ BU , sS cj , t s.t. constraints (18a-18b),(19-26),(27a-27b),(28a-28d),(31a-31c),(32a-32d), (33-34),(43),(45),(48a-48c), and all the constraints corresponding to production units scheduling operations in model (IP) (P2) The restricted relaxed models (P1-R) and (P2-R) are constructed by fixing binary variables using equations (37a)-(37h) in models (P1) and (P2), respectively. Infeasible restricted relaxed model is obtained by removing equations (48a)-(48c) from models (P1-R) and (P2-R). The upper bounding problem (IP-U) is obtained by fixing values of certain variables using equations (38a)-(38d) in full-scale model (IP) presented in chapter 4. The models (PC-s) and (PC-j) are obtained from (P2) by replacing its objective function by (39) and (40), respectively. 5.6.1. Updating the Multipliers The subgradient method first proposed by Fisher (1981) is commonly used method to update Lagrangian multipliers used in solving Lagrangian relaxation problems. It requires solving all sub-problems at each iteration to find a subgradient of the relaxed problem to compose a search direction to update the multipliers. The elements of subgradient to the relaxed problems are as follows: B P g Tsj ,t Ts j ,t Ts j ,t , j J BU , t T g Tfj ,t Tf B j ,t Tf B P j ,t , j J BU , t T P g sCf, j ,t Cf s , j ,t Cf s , j ,t , j J BU , s S cj , t T 135 Cf Where B B s , j ,t , Ts j ,t , Tf B j ,t and Cf P s , j ,t P , Ts j ,t , Tf P j ,t are the values of the duplicating variables, obtained from the solution of the blend scheduling problem and production scheduling problem, respectively. The multiplier is updated for the next iteration according to u( m1) u( m) s( m) d ( m) Here, u( m) denotes the multiplier values, s( m) is the step size, and d ( m) is the search direction in iteration m. The search direction is usually set to the subgradient, but it is more efficient to use modified formula that takes into consideration pervious iteration direction. Search direction is defined as recursive formula as suggested in Gaivoronski (1988), where d (1) g (1) : d (m) g (m) d ( m 1) 1 , m 1 Where determines how much consideration is given to the previous direction, 0 being no consideration to previous direction and 1 being an average of current and previous search direction. The step size s( m) is given by a widely used formula: s (m) (m) Z U L(u ( m ) ) g (m) 2 Here Z U is the upper bound of (IP) , L u ( m ) is the solution of the relaxed problem at u( m) , and ( m) should be assigned a value in the interval ( , 2 ), where 0 to ensure convergence. If there is no improvement in the lower bound in K successive iterations, we set ( m) 0.5 ( m1) . ( m) is reset back to (0) whenever an improved upper bound or lower bound. 5.6.2. Stopping Criteria In this section, we present the stopping criteria used to terminate the Lagrangian algorithm. - Stop if the feasible solution is found at lower bound, which happens when g ( m) 0 ; but this rarely occurs for large problems. 136 - Stop if the lower bound from the Lagrangian relaxation exceeds the best known upper bound, i.e., if Z L Z U . - Stop if the duality gap is less than or equal to 1%. ZU Z L duality gap L Z Stop if d ( m ) , s( m) , or Z U Z L . - In practice we do not wait till above mentioned stopping criteria is met, so we impose the number of iteration as the stopping criterion, i.e., stop after a limited number of iterations. 5.6.3. Algorithm Here using u( m) at mth iteration, the objective value of the relaxed problem and restricted relaxed problem is calculated from equation in (4b) and given as Z R u ( m ) and Z RR u ( m ) respectively. Z ( m) is the objective value of the feasible solution through constructed heuristic at the mth iteration. The procedure for the restricted-LD algorithm: Step 1: Solve pre-processing problems (PC-j) and (PC-s) as described in section 5.5.1 Step 2: Initialize m=0; Z U ; Z L ; uTsj ,t (0) 0; u Tfj ,t (0) 0; usCf, j ,t (0) 0; (0) ; Step 3: Solve relaxed problem by decomposition that is solve independent problems (P1) and (P2). B Obtain Z R , Cf s , j ,t , violatn R R total , and chgovrR . Step 4: Construct a restricted relaxed problem as described in section 5.4.2. Solve the restricted sub-problems (P1-R) and (P2-R). If infeasible, then solve corresponding infeasible sub-problem using model (P1-RI) or (P2RI). 137 B Obtain Z RR , Cf s , j ,t RR RR , violatntotal , and chgovrRR . RR R Step 5: If violatntotal and Z RR u ( m ) Z L , then Z L Z RR u ( m ) . violatntotal RR R Otherwise, if violatntotal and Z R u ( m ) Z L , then Z L Z R u ( m ) . violatntotal Step 6: Construct a feasible solution to the original problem as described in section 5.4.2.2. U If Z m Z U , then Z U Z m and obtain parameters violatnUdem , violatntotal , and chgovrU for constraints described in section 5.5.3. Step 7: Update Lagrangian multipliers by using the sub-gradient method presented in section 5.6.1. Step 8: Terminate the algorithm if the current solution satisfies at least one “stopping criteria” listed in section 5.6.2. is met. Otherwise, set m m 1 and return to step 3. In step 7, the Lagrangian multipliers are calculated using restricted relaxed problem solution RR R unless violatntotal . Problems in step 1 are independent and can be solved in parallel. violatntotal Similarly, optimization models (P1) and (P2) in step 3 and models (P1-R) and (P2-R) in step 4 are independent of each other and can be solve in parallel to reduce computational time. The flowchart of the framework for the algorithm is given Figure 5.5. 138 Solve (PC-j) and (PC-s) Initialize , , and Lagrange multipliers Solve (P1) and (P2) Solve (P1-R) and (P2-R) Solve corresponding infeasible problem No Sub-problems feasible Yes No Yes If Yes , , If Done If Solve IP-U , Terminating Criteria No Update multipliers Figure 5.5. Algorithm for the proposed restricted-Lagrangian decomposition (restricted-LD) 139 5.7. Computational results The computational results are obtained on Dell Precision (IntelR XeonTM with CPU 3.20 GHz, 3.19 GHz, and 2 GB memory) running on Linux using CPLEX 12.3.0/GAMS 23.7.2 to demonstrate their effectiveness in solving oil-refinery scheduling problems using 6 different examples, where each example differs in either demands values, intermediate due dates, and initial hold up in the tank. The Lagrangian algorithm is set up in MATLAB and interfaces with GAMS to solve scheduling problems. The algorithm is evaluated with three performance measures: duality gap, number of iterations, and computational time (seconds). The maximum solution time of 28,800 CPU seconds is used for full-scale problem and for the decomposed problem maximum solution time of 3,600 CPU seconds is used. Optimality tolerance of 1e-6 is used as termination criteria for all examples solved using CPLEX. For the algorithm, the maximum number of iteration is set to 100, 16 , and K = 2. Table 5.1 Model statistics for preprocessing problems (PC-j) (PC-s) Diesel Blender Jet Blender Event pt. 3 3 3 Int./Cont. var. 120/924 120/980 120/978 (Constraints) (2425) (2425) (2425) Nonzero Elem. 7753 7813 7813 The preprocessing problems for each blend components and blend units are solved before Lagrangian algorithm is implemented. The model statistics for preprocessing models are presented in Table 5.1 and these models are solved in less than 1 CPU seconds. The model statistics for full-scale integrated model and relaxed sub-problems for 6 different examples are shown in Table 5.2. Total number variables in restricted relaxed sub-problems, (P1-R) and (P2- 140 R), are same as that of corresponding relaxed models because (P1-R) and (P2-R) are obtained by restricting solution space by including additional constraints and fixing values of certain binaries/continuous 0-1 variables in corresponding relaxed models, (P1) and (P2). Table 5.2 Model statistics Blend scheduling model (P1) Full Scale Production scheduling Event pt. Model model (P2) Int./Cont. variables Ex # Event pt. Event pt. (Constraints) . Orders Int./Cont. var. Int./Cont. var. Nonzero Elem. (Constraints) (Constraints) Diesel Nonzero Elem. Jet Blender Nonzero Elem. Blender 1 2 3 4 5 5 5 5 360/2663 127/958 33/247 200/1526 (6387) (2489) (717) (4649) 22633 9819 2323 15394 5 5 5 5 360/2663 127/958 33/247 200/1526 (6379) (2481) (717) (4649) 22569 9760 2329 15405 5 5 5 5 360/2663 127/958 33/247 200/1526 (6379) (2480) (717) (4649) 22569 9725 2324 15395 5 5 5 5 380/2771 139/1022 41/291 200/1526 4 4 4 6 141 5 6 (6741) (2695) (847) (4648) 24018 10598 2794 15399 6 6 6 6 484/3453 184/1309 60/402 240/1835 (8644) (3649) (1272) (5794) 31765 14956 4384 19442 7 7 7 7 736/5005 277/1829 83/530 280/2146 (13394) (5585) (1833) (7027) 53145 23804 6643 23903 8 13 The computational results for full-scale model and Lagrangian algorithm are shown in Table 5.3. Proposed restricted-Lagrangian decomposition algorithm is quite effective for these refinery scheduling problems and we obtain good solutions to the scheduling problem. Figure 5.6 to Figure 5.11 shows the convergence of upper and lower bound using the classical LD and restricted LD algorithms in examples 1-6, respectively. The proposed restricted-LD algorithm outperforms classical Lagrangian decomposition in terms of quality of the solution, duality gap, number of iterations, and computational time. In classical approach, the lower bound is obtained solely from relaxed problem‟s solution, whereas proposed restricted-LD picks the best lower bound between the solution of relaxed problem and restricted relaxed problem. Thus, restrictedLD lower bounds take into consideration the continuous process characteristic of the refinery units and gives better lower bound as observed in Figure 5.6 to Figure 5.11. In most cases, both algorithms provide an upper bound closer to the optimal solution of original problem in first iteration because feasible problem is constructed using restricted relaxed problem solutions and has flexibility of obtaining a better solution because only the certain binaries are fixed to 1 and none are fixed to 0. In situation where first iteration doesn‟t provide an 142 upper bound closer to optimal solution, the upper bound improves vastly early on and ultimately provides a better solution. In example 5, the first iteration provides an upper bound that is very far from an optimal solution, however, it improves greatly in second iteration, and by fifth iterations it‟s almost closed to the optimal value. After finding the best upper bound early on, both algorithms spend rest of the computational time proving optimality. Table 5.3 Computational Results Full Scale Model E x Gap CPU z Classical LD Restricted LD 1.50, 0.8 1.20, 0.7 Gap CPU Total z (%) sec Gap CPU Total (%) sec Iter z (%) sec Iter 1 450.77 0.00 342 450.77 29.34 652 43 450.77 -0.21 113 11 2 598.64 0.00 367 598.64 5.92 2133 57 622.91 0.15 251 9 3 6222.54 0.00 8952 6222.54 1.49 15304 71 6222.54 0.79 3245 14 4 592.81 0.00 1535 602.33 25.22 1048 26 592.81 0.95 1225 40 5 503.19 0.00 6458 504.25 12.62 1234 22 503.86 7.72 2652 20 6 749.35 72.8 28800 552.50 18.04 36865 20 512.65 -0.86 20490 15 The performance (duality gap) of restricted-LD algorithm is better than that of classical approach because restricted-LD is able to improve lower bound faster than classical LD. Proposed restricted-LD provides better duality gap than classical LD and restricted-LD is terminated in almost all cases when gap is less than 1% as shown in Table 5.3. In case of classical decomposition approach, the algorithm was terminated for all examples when the step size meets the predefined tolerance of 1e-6. The step size reaches the tolerance faster because we cut the step size parameter ( m) by half if the lower bound does not improve after 2 iterations. Even when the duality gap is high, the upper bound is closer to the original problem optimal solution in both algorithms. 143 Table 5.4 Time spent (%) in each step compared to the total solution time of Lagrangian decomposition algorithms Problems Relaxed Ex. 1 Ex. 2 Ex. 3 Ex. 4 Ex. 5 Ex. 6 77.32 91.47 98.28 90.78 94.16 93.31 Classical Restricted relaxed 11.48 4.40 0.96 4.17 2.80 2.95 4.13 0.75 5.05 3.04 3.74 LD Feasible 11.20 Relaxed 68.12 87.92 98.76 89.61 97.48 91.96 Restricted Restricted relaxed 13.26 4.94 0.49 4.19 1.31 7.96 7.14 0.75 6.19 1.21 0.076 LD Feasible 18.63 Table 5.4 compares the computational effort exerted to solve relaxed problem, restricted relaxed problem, and feasible upper bounding problem. As expected, majority of the time is spent in solving relaxed problem. Classical LD takes 652 CPU seconds to solve example 1 and spends 77.32% of the total time on solving for relaxed problem and only 11.48% on solving restricted relaxed problem. Similarly for restricted-LD, example 1 takes 68.12% of the total computational time on solving relaxed problem, 13.26% on solving restricted relaxed problem, and 18.63% on solving upper bounding problem MILP problem. Even for the large scale complex problem given in example 6, the upper bounding MILP problem can be easily solved. Computational time can be further improved by solving relaxed sub-problems in parallel and similarly restricted relaxed problems can be solved in parallel. 144 Figure 5.6 Convergence of upper and lower bound of Lagrangian decomposition for example 1. 145 Figure 5.7 Convergence of upper and lower bound of Lagrangian decomposition for example 2. 146 Figure 5.8 Convergence of upper and lower bound of Lagrangian decomposition for example 3. 147 Figure 5.9 Convergence of upper and lower bound of Lagrangian decomposition for example 4. 148 Figure 5.10 Convergence of upper and lower bound of Lagrangian decomposition for example 5. 149 Figure 5.11 Convergence of upper and lower bound of Lagrangian decomposition for example 6. 5.8. Summary This chapter introduces a Lagrangian decomposition framework, restricted Lagrangian relaxation problems for integrated production unit scheduling and finished product blending and delivery scheduling problems. The algorithm is built based on the mathematical formulation given in Chapter 4. A restricted relaxed Lagrangian problem is constructed to facilitate generation of tighter upper bound in each iteration. Restricted relaxed sub-problems are based upon the solution of Lagrangian relaxed sub-problems and solution of restricted relaxed problem takes into consideration the continuous process characteristic of rundown streams. To improve the performance of the algorithm, a preprocessing step, constraints for decomposed sub-problems, and inclusion of the best upper bound‟s solution in lower bounding problems are proposed. The 150 proposed restricted relaxed sub-problems produce better lower bounds and better upper bounds. Computational results of a real case study show that the proposed algorithm is very effective and provide better solutions in reasonable times. Nomenclature Indices i Tasks j Production units k Storage tanks n Event-points m Lagrangian algorithm iteration counter o Product order p Properties s States t Continuous time-points similar to event-points Sets Ij Tasks which can be performed in unit j I sc / I sp Tasks which can consume/produce material s J All Production units J BU Blend units J PU Production units that belong to production unit operations J sc / J sp Units that consume/produce material s Ji Units which are suitable for performing task i Jh Units that can produce all the same products as some other unit in the 151 refinery Jm Units which are suitable for performing multiple tasks J seq j Units that follow unit j (no storage in between) J kkp / J kpk Units that consume/produce material s stored in tank k JP Units that produce products K Storage tanks Kh tanks that can store the same products as some other tank in the refinery pk K kp j / Kj Tanks that store material consumed/produce by unit j Km Multipurpose tanks that can store multiple materials Kp Tanks that can store final products Ks Tanks that can store material s N Event-point within the time horizon O Orders for products that are stored in tanks Os Orders for finished product s P Product Properties S States SA Group A products, stored in tanks SB Group B products, not stored in tanks Sbc States that are blended by blend-units Sk States that can be stored in tank k Sic / Sip States that can be consumed/produced by task i S cj / S jp States that can be consumed/produced by unit j Parameters 152 Do, s , Do, s Demand limit requirement for order o and product s that is stored in tank max_rates Maximum rate of production of material s oto,i 1 if task i is performed when order o of PUSP is processed by BSP. rs Demand of the final product s at the end of the time horizon RBmin / RBmax j j minimum and maximum production rate of a blend unit j RPjmax Maximum rate material is supplied by (PSP) to blend unit j RPsmax Maximum rate blend component s is supplied by (PSP) to blend unit j max Rimin , j / Ri , j Minimum/ maximum rate of material be processed by task i in unit j RLi Minimum run length for task i RU kmin / RU kmax Minimum/maximum rate of product unloading at tank k stos , k Amount of state s that is present at the beginning of the time horizon in k tak Fill draw delay for product tank k UH Available time horizon Vkmax Maximum available storage capacity of storage tank k Vkheel Maximum heel available for storage tank k yos , k 1 if the material s is present at the beginning of the time horizon in k max smin ,i / s ,i Proportion of state s produced/consumed by task i max bsmin , j / bs , j Minimum/maximum proportion of blend component s consumed by blend unit j Variables Binary Variables wvi , j , n Assignment of task i in unit j at event-point n ins , j , k , n Assigns the material flow of s into storage tank k from unit j at point n 153 Assigns the starting of product flow out of product tank k to satisfy order o lk ,o, n at event-point n outs , k , j , n Assigns the material flow of s out of storage tank k into unit j at point n ys , k , n Denotes that material s is stored in tank k at event-point n 0-1 continuous variables j ,n For unit j, 1 if the unit becomes active for very first time at event-point n k , n For tank k, 1 if the tank becomes active for very first time at event-point n s , s , k , n 1 if material in tank k switchover service from s at event-point n to s‟ at later event-point os , s, k 1 if material in tank k switchover service from s to s‟ i , i , j , n 1 if task at unit j changes from i at event-point n to i’at later event-point. a j ,n Assigns the materials flow to blend unit j at event-point n x j , n ,t Denotes that blend unit j is active at event-point n and time-point t Positive variables bps ,i , j , n Amount of material s produced task i in unit j at event-point n bcs ,i , j , n Amount of material s undertaking task i in unit j at event-point n Amount of blend component s consumed/supplied at blend unit j at time- Cf sB, j ,t / Cf sP, j ,t point t dgol Minimum demand quantity give-away term for order o dgou Maximum demand quantity give-away term for order o H Total time horizon used for production tasks JJf s , j , j , n Flow of state s from unit j to consecutive unit j’ for consumption at point n Kif s , j ,k ,n Flow of material s from unit j to storage tank k event-point n 154 Kof s , k , j , n Flow of material s from storage tank k to unit j at point n Lfo, k , n Flow of final product for order o from storage tank k at event-point n mhs , k , n Maximum heel give-away term for product tanks rgs Minimum demand quantity give-away term for Group B product s Rif s , k , n Flow of raw material to storage tank k event-point n sts , k , n Amount of state s present in storage tank k at event-point n Amount of state s that is downgraded to state s‟ in storage tank k at event- std s , s, k , n point n Tearlyo Early fulfillment of order o than required Tfi , j , n Time that task i finishes in unit j at event-point n Tf jB,t / Tf jP,t Finish time for materials consumption/supply at blend unit j at time-point t Tlateo Late fulfillment of order o than required Tosk ,o, n Time that material starts to flow from tank k for order o at event-point n Tof k ,o, n Time that material finishes to flow from tank k for order o at event-point n Tsi , j , n Time that task i starts in unit j at event-point n Ts Bj ,t / Ts Pj ,t Start time for materials consumption/supply at blend unit j at time-point t Tsf j , k , n Time that material finishes to flow from unit j to tank k at event-point n Tsf k , j , n Time that material finishes to flow from tank k to unit j at event-point n Tss j , k , n Time that material starts to flow from unit j to storage tank k Tssk , j , n Time that material starts to flow from tank k to unit j at event-point n Uif s , j ,n Flow of raw material s to production unit j at point n Uof s , j , n Flow of product material s from unit j at point n UUf j , s ,v , n Amount of state s received by unit j at period v is processed at event n 155 z Objective value of full-scale integrated model 156 Chapter 6 6. Efficient Decomposition Approach for Large Scale Refinery Scheduling This chapter proposes a heuristic algorithm for integrated scheduling of refinery production unit operations and finished product blending and delivery operations. This work addresses the interdependences of the production unit operations and finished product blending and delivery operations in the refinery that has components streams coming directly from process units to blend headers with no components storage options available. To apply the proposed iterative algorithm, the integrated oil-refinery scheduling problem is decomposed into production unit scheduling problem and into finished product blending and delivery scheduling problem. The goal of the algorithm is to obtain a feasible solution that satisfies products demands commitments with minimum demurrage violations while minimizing inventory cost and penalties subject to product and logistics giveaways. The applicability of the algorithm is illustrated by a realistic large-scale refinery case study producing diesel and jet fuel. Significant savings are realized in the computational effort. 6.1. Introduction Even with inclusion of valid inequalities the refinery operations scheduling model is still computationally prohibitive. The mathematical decomposition approach presented in previous chapter can lead to global optimal solution, however for practical application where the need for feasible solution is greater than obtaining the best solution possible, the solution times of mathematical decomposition is undesirable. For this situation, heuristics approach exploiting inherent structure of the problem alongside insight into the makeup of a feasible solution provides an alternative to massive full-scale optimization problem or over mathematical decomposition. Furthermore, heuristics approach can be used as preprocessing step for mathematical decomposition solution strategy to reduce complexity of the original model (i.e. reduce feasible 157 solution space). (Amaldi et al., 2008; Arief et al., 2007; Cornillier et al., 2008; Karuppiah et al., 2010; Kelly, 2003b; Roslof et al., 2002) Usually blend component tanks are present when the blending and delivery sub-problem is addressed without simultaneously addressing production unit scheduling problem. However, when blend components tanks are not available, blend units are directly connected to the upstream production units. In order to comply with the finished blend product properties specifications, blend recipe needs be determined by considering the production constraints of upstream processes so that the blend components production and consumption rate is consistent across two scheduling decision levels. These interdependences of the blending operations and production unit scheduling operations require an integrated approach to the refinery scheduling. In this work, we decomposed the large-scale integrated problem into a production unit scheduling problem and a finished product blending and delivery problem. A heuristic algorithm is proposed based on the decomposed network to obtain a good solution of the original integrated problem in reasonable computational times. The algorithm is built upon the mathematical formulation developed given in Chapter 4. The proposed decomposition approach focuses on effectively solving the decomposed problem to obtain a feasible solution that satisfies demands while minimizing due date violations, product giveaway, task changeovers at multipurpose blend units and tanks, and inventory costs. The outline of the chapter is as follows. Section 6.2 presents additional valid inequalities, section 6.3 presents proposed solution approach and the heuristic algorithm is discussed in section 6.4. A realistic case study presented in section 4.2.1 is used to illustrate the applicability of the proposed algorithm to a large-scale model in section 6.6. 6.2. Mathematical Formulations Mathematical formulations proposed earlier for the oil-refinery operations scheduling problem is based on the continuous time representation and an idea of unit-specific event points. (M. G. 158 Ierapetritou & C. A. Floudas, 1998; M.G. Ierapetritou & C.A. Floudas, 1998; Ierapetritou et al., 1999) State-task network (STN) representation introduced by Kondili et al. (1993) is used in formulating the problem. Event based formulation enables the representation of time in a continuous manner without unnecessary time slots or intervals. Although the location of the event points is unknown, the number of event points has to be considered initially. It is not straightforward to know the minimum number of event points needed to achieve the global optimum solution. Choosing a large number of event points makes it more likely to achieve the global optimum solution but it also results in larger solution times. A number of advances of the original methodology have been proposed in the literature to improve the scheduling formulation. (Janak et al., 2004; Mouret et al., 2011) The model involves material balance constraints, capacity constraints, demand constraints, quality constraints, logistics constraints, and setup constraints. The demand constraints ensure that all the products demand are satisfied while the quality constraints ensures that product quality specifications are met. To preserve the linearity of the formulation, linear blending rules are considered. The logistics constraints include all the operational rules presented in the previous section and the setup constraints models the activation of parallel production units and parallel storage tanks during the scheduling horizon. Due to production capacity limitation, downgrading of a higher quality product to a lower quality product may be necessary to satisfy demand on time and constraints are added in the model to formulate product downgrading in multipurpose storage tanks. If a feasible solution that satisfies all the quantity and logistics constraints cannot be obtained, then it is essential to produce a schedule that can still be implemented in real-life oil-refinery sacrificing model feasibility. In this case, we introduced artificial variables to treat any infeasibility present and these variables are subsequently penalized in the objective function to obtain an optimal solution that satisfies as many as possible from the quantity logistics constraints by minimizing giveaways. The objective function (1) is used to maximize the performance, minimize inventory costs and product demands and intermediate due-dates violations. 159 z i , jJ i , n ci1, j wvi , j , n k K p , o , n ci6,i i ,i , j , n jJ , iI j , i I j , i i , n m c sS B s , k K s ck3 sts , k , N c 4j j , n j J h , n c std s , s , k , n 9 k k K m , sS k , n ck5 k , n k K h , n cs7, s os , s , k k K , sS k , s S k , s s m k K m , sSk , s S k , s s , n 13 s ck2 lk ,o , n cs8, s s , s , k , n k K , sS k , s S k , s s , n m c mhs , k , n co11dg ol co12 dg ou 10 s,k oO (1) oO rg s c Tearlyo c Tlateo oO 14 o oO 15 o The oil-refinery performance is represented by longer run-modes ( wvi , j ,n , lk ,o,n ) and minimization of inventory ( sts ,k ,n ), start up set-ups ( j , n , k ,n ), changeovers ( i ,i, j ,n ,os, s, k ,s, s, k , n ), maximum heel violations ( mhs ,k ,n ) and production downgrading ( std s , s,k ,n ). Logistics giveaways, under and over production, and demurrage are penalized. The penalty weights are assigned arbitrarily to each term depending on its importance in schedule. The deviation from intermediate due dates ( Tearlyo , Tlateo ) and under/over production of finished products ( dgol , dgou , rgs ) are heavily penalized. It is favorable that the multipurpose product tanks store different grade products only in a certain order that is allowed by the sequence dependent switchovers constraint. The favorable switchovers are from higher grade of product to lower grade and unfavorable switchovers are from lower grade to higher grade product. Due to contamination issues, unfavorable switchovers are more heavily penalized than favorable. Similar to multipurpose tanks, there is a sequence dependent switchover restriction for multipurpose blend units. Note that the different penalty parameters have significant effect on the computational time required to obtain an optimal solution. All these constraints and objective function give rise to a large scale, incomprehensible mixed integer linear programming (MILP) model that becomes computationally expensive to solve using standard optimization software to optimality. Valid Inequalities 160 To improve the computational efficiency, valid inequalities for the refinery scheduling model are introduced in previous chapter. These valid inequalities are based on certain logistics and operational rules and material balance. These inequalities are redundant to the original model because the optimizer will eventually discards infeasible solution and arrive at a feasible solution. Thus, these inequalities do not cut feasible solutions space but rather eliminate the region of infeasible solution explicitly instead of waiting for optimizer to reach that conclusion. Application of valid inequalities in scheduling model to speed up convergence to optimal solution can be found in work of Velez et al. (2013), G. K. D. Saharidis and Ierapetritou (2009), and Khor et al. (2012). The valid inequalities that are included in the full-scale integrated model used in this work are as follows: Constraint (1a) connects total amount of material consumed and produced at unit j and event point n in a form of inequality since material balance constraints present in the fullscale model uphold material balance requirements. Kof s , k , j , n JJf s , j , j , n Uif s , j , n p j J seq j Js k K kp j Ks k K jpk Kif s , j , k , n JJf s , j , j , n Uof s , j , n , j J , n N (1a) c j J seq j Js Ks Constraints (1b-1d) reaffirms binary variable ys,k ,n to be 1 if there is material s flowing in and out of the storage tanks k at given event point n. in s, j ,k ,n jJ kpk out jJ kkp ys , k , n , k K , s S k , n N jJ kpk s, j ,k ,n jJ kkp ys , k , n , k K , s S k , n N (1b) (1c) 161 l oOs ys , k , n , k K bp , s Sk , n N k ,o, n (1d) If the tank is sending or receiving material s from a production unit j at event point n, then that unit will be active at event point n and these allocation requirements are represented by equations (1e-1f). ins , j , k , n outs , k , j , n k K jpk Ks k K jpk wvi , j , n , j J pk , s S jp , n N (1e) wvi , j , n , j J kp , s S cj , n N (1f) K s iI j I sp c k K kp j K s iI j I s k K kp j Ks Constraints (1g-1h) asserts material s flow in/out of tanks to be active at event point n if the unit j is processing the material at that event point. ins , j , k , n outs , k , j , n k K jpk wvi , j , n , j J pk , s S jp , n N (1g) wvi , j , n , j J kp , s S cj , n N (1h) iI j I sp Ks iI j I sc k K kp j Ks In addition to the valid inequalities proposed in our earlier work, we propose here some new valid inequalities to further improve the computational performance. If two units are consecutive without any storage tank between them, then constraint (1i) affirms the simultaneous operation of these units due to the continuous operation mode. However, this constraint is not imposed on parallel production units that can produce the same type of products. For units that follow or are followed by parallel units, valid inequalities (1j-1l) are included. wv iI j i, j ,n wv iI j i , j , n , j J / J h , j J / J h , j j , j J seq j ,n N (1i) 162 wv wv wv iI j iI j iI j i, j ,n wvi , j , n , j J h , n N , and j J h , j J seq j (1j) wvi , j , n , j J / J h , n N , and j J h , j J seq j (1k) j J h , jJ seq j , iI j i, j ,n j J h , jJ seq j , iI j i, j ,n j J h wvi , j , n , j J / J h , n N , J seq Jh j (1l) J seq j , iI j As mentioned earlier, the demand order set O is arranged according to the ascending due date start time. That is, timeso timeso , o, o o, o o . Operational requirement for the finished blend product tank is that loading and unloading cannot happen at the same time, thus if the material is unloaded at tank k at event point n+1 then the tank should not be empty at previous event point n . This is condition is captured by equation (1m). l k , o , n 1 o y sSk s,k ,n , k K bp , n N , n N (1m) The demand order set O is arranged in chronologically ascending order based on intermediate due-dates. We apply inequality (1n) to require that the demand order o to be satisfied by product tank k earlier than order o o . Furthermore, if the initial inventory of the products is less than the required total minimum demand orders, then the production should take place before the demand is fulfilled. This requirement is upheld by constraint (1o) where prs lk , o , n k K s , n n lk , o , n lk ,o , n , k Kbp , s Sk , o Os , o Os , o o, n N iI sp , jJ i , n n wvi , j , n , s S A Sbp , k K s , o Os , prs 0, n N , n 1 D oOs o, s sto k K s s,k . (1n) (1o) The mathematical formulation presented in presented in last chapter, inequalities (1a-1o), and objective function (1) comprise the large-scale MILP model for an integrated scheduling problem. 163 6.3. Solution Strategy In this section, we propose a heuristics-based algorithm for solving full-scale integrated problem within reasonable computational times. The algorithm involves solving independent subproblems in an iterative fashion until a feasible solution is found. Spatial decomposition is applied to refinery system under study as shown in Figure 6.1, to obtain smaller sub-problems that are easier to solve than full-scale integrated problem. (Jia & Ierapetritou, 2003; Karuppiah et al., 2008) The refinery system under study involves rundown blending operations, thus overall refinery structure is decoupled by splitting the blend components stream originating from process units operations and supplying directly to blend headers. This decoupling produces two scheduling sub-problems relating to the production unit operations and blending and delivery operations of finished products, respectively. The complicating variables associated with split pipelines are defined differently in decomposed scheduling sub-problems than are in the integrated problem. Production unit scheduling problem defines variables associated with component streams as demand variables whereas blend scheduling operations defines the component flow variables as raw materials variables, as seen in Figure 6.1. Hence, the main goal of the production unit scheduling sub problem (PSP) is to satisfy the demand requirement of final products that belong to set S B and demand of the blend components required by the blend scheduling sub problem. Similarly, the purpose of the blend scheduling sub problem (BSP) is to satisfy the demand of the finished blend products by mixing the raw materials (blend components), supplied by PSP, following the blend recipe and product property specifications. 164 Intermediate Products Finished Blend Products Production CDUs units Storage and other Production units Blend headers Production Unit Operations Shipping/lifting points Product tanks Product Blending and Delivery Figure 6.1 Decomposition of an oil-refinery operations network by splitting blend components pipelines Dummy Tanks for Blend Components Production Scheduling Problem …. …... …... …... Production Unit Operations Blending and Delivery Operations Blend Scheduling Problem Figure 6.2 Illustration of scheduling sub-problems obtained by splitting component flow streams connecting process units directly to blend headers In refinery system under study, blend components flow from process units to blend headers happen simultaneously with blending operations. Thus, in decomposed system, production unit 165 scheduling problem needs to satisfy demand of blend components within certain time frame. We introduce dummy tanks for blend components in PSP as seen in Figure 6.2 for ease of imposing strict intermediate due-dates for components demand. These dummy tanks acts as product storage tanks in PSP but have different operational rules than finished blend product tanks. These operational rules as follow: - Each dummy tank are dedicated to a blend component - Simultaneous loading and unloading of blend components - No accumulation of component at the end of an event-point - Dummy tanks can supply components to multiple blend headers at same time - Multiple tanks supplying directly to the same blend header must do so at the same time Constraints relating to these rules are presented in detail later, but first, a general outline of the proposed iterative algorithm is presented in Figure 6.3. The algorithm solves BSP, PSP and restricted BSP in iterative fashioned until a global feasible solution that satisfies the final products demand requirements while minimizing demurrage. Here, the restricted blend scheduling problem (RBSP) is derived by limiting the feasible solution space of BSP based on solutions of PSP and BSP. Blend Scheduling Problem (BSP) Demand targets for Blend components Production Unit Problem (PSP) Supply of Blend components Restricted Blend Scheduling Problem (RBSP) Integer Cuts 166 Figure 6.3 Proposed solution approach. The details of BSP, PSP, and RBSP are presented in the following sub-sections. These problems include all the constraints present in the full-scale model and some additional constraints that are necessary to facilitate the convergence to a global feasible solution at the end of iteration. The demand constraint is given in equation (2) and a feasible global solution of the heuristic algorithm must satisfy the minimum demand requirements hence demand give-away variables dgol 0 and rgs 0 in equation (2). Do, s rs dg ol rg s Lf k ,o , n Uof s , j , n Do, s dgou , s Sb , o Os s S f k ,n j ,n (2) 6.3.1. Blend Scheduling Problem Blend scheduling problem (BSP) involves blend headers, finished blend products storage tanks, delivery of these products according to intermediate due-dates and raw materials being supplied directly to blend headers. Refinery may include more than one blend headers producing multiple blend products. In the event that refinery has more than one blend header, we can further decompose BSP into smaller scheduling sub-problems depending upon the interactions between blend headers. The blend scheduling sub-problems must not share common blend components and finished product tanks and should produce different types of finished blend products. Hence, blenders using shared components must be jointly optimized. Decomposed independent blend scheduling sub-problems can be solved in parallel. The mathematical model for BSP contains all the constraints that are present in the original model presented in chapter 4 and some additional constraints. To obtain a solution that provides minimum amount of blend component necessary for satisfying finished product demands, we fix the maximum finished blend-product demand to the minimum demand 167 requirement, that is, Do, s Do, s . Furthermore, the demand give-away variables are fixed to zero, dgol 0, dgou 0 . Thus, the demand constraint is rewritten as follow for BSP: k K s , n Lf k ,o , n Do, s , s S A , o Os (3a) Where, parameter Do, s .is a minimum demand of product s and order o and variable Lf k ,o,n is the flow of product from product tank k to satisfy to order o at event-point n. Furthermore, instead of having amount of product satisfied by tank k to be bounded by minimum and maximum unloading rate, RU kmin and RU kmax , as is in full-scale model; for BSP, unloading rate is fixed to the maximum and the capacity constraints given in original model can be restated as follow: Lf k ,o , n RU kmax Tof k ,o , n Tosk ,o , n , k K p , o O, n N (3b) To consider the production capacity of the process units producing blend components, maximum blend components flow rate constraint (3c) is included. The maximum blend components flow rate ( max_rates ) parameter is determined a priori using the maximum production recipe and maximum production capacity of the process units. Uif s , j , n max_rates Tf i , j , n Tsi , j , n UH 1 wvi , j , n , j J b , i I j , s Sic , n N (3c) Furthermore, we can reduce the feasible solution space by not allowing the production to take place at the very last event point since the blend scheduling problem must respect the loading and unloading restrictions of product tanks, which means that material produced at the last event would not contribute to satisfying the demand and would just accumulate. wvi , j ,n 0, j J b , i I j , n N , n N (3d) 168 Restriction given in equation (3d) reduces the computational cost involve solving BSP without sacrificing an optimal solution of BSP. 6.3.1.1. Acquiring Parameters for PSP and RBSP from Solution of BSP The optimal solution of the BSP provides the information regarding the amount ( Uif s , j ,n ) of blend component s flowing into each blend header j at event point n to produce blend product s using task i. The start and finish time ( Tsi , j ,n and Tfi , j ,n ) of the blend component flow into the unit is also available. Based on this information, we derive demand, due dates, and other parameters for the production unit scheduling problem. Each blend header gives rise to Sbcj n j demand orders o O j that needs to be satisfied by PSP. Here, Sbcj S cj Sbc and Sbcj is the cardinality of the set Sbcj , n j wvi , j , n , and O j is a set i,n of demand orders comprising of blend components consumed by blend unit j. Parameter MN max n j 1 . The parameter oto,i is 1 if task i is performed by a blend header j when order o j is processed. A parameter ato 1 for the demand orders determined using the solution of BSP. The active event points of each blend unit j are used to assign values to demand orders as: Do, s Uf s , j ,n , Do, s Uf s , j ,n , timeso Tsi , j , n , timefo Tfi , j , n , and oto,i wvi , j ,n . That is, if a given blend header is active for 3 event points and can blend 3 components, a total of 9 active demand orders correspond to the blend header, as seen in Figure 6.4. In order to take into consideration the interactions of the production units in PSP and to allow flexibility for over production, Sbcj additional demand orders are included to the set O j . These additional demand orders are also necessary to satisfy group B product demands and to meet blend component due-dates without causing infeasibility in PSP. These demand orders have values of Do, s 0 , Do, s M , timeso 0 , timefo UH . Parameter oto,i for these additional demand 169 orders is equal to the task mode of the last active event point ( n j ) so that the task changeovers cost is minimized. Each demand order only corresponds to one blend component, thus the cardinality of the order set O for PSP is j J Sbcj n j 1 . Demand orders greater than Sbcj n j do B not need to be satisfied by PSP and are called non-active demand orders and ato 0 . In Figure 6.4, demand orders o10, o11, and o12 are non-active demand orders. n1, i2 o1 o2 n2, i3 o3 o4 o5 n4, i1 o6 o7 o8 o9 o10 o11 o12 Figure 6.4 Determining parameters for PSP using a solution of a blend header j Flow into a product tank k Flow out of a tank k x y Figure 6.5 Determining parameter tupo for PSP using the corresponding blend header j The in-line blender mixes multiple components at the same time to produce final products, thus we introduce a parameter bo j ,o to capture the relation between different demand orders for a blend unit j. As mentioned before, each blend unit has Sbcj n j 1 demand orders belonging to set O j . We assign the value of 1 to the first set of demand orders that are processed at the same time by the blend unit j. For the last set of demand orders, bo j ,o n j 1 . Note that the 170 set O j is arranged in an ascending chronological order based on due-date. Figure 6.4 shows how active demand orders are assigned and how parameter bo j ,o is determined. To minimize demurrage violations for finished blend-products in the optimal solution of RBSP, a feasible solution of PSP should satisfy blend components demand within certain time frame. For this purpose, a parameter tupo is determined from the optimal solution of BSP. A product tank k receives finished product produced by blend unit j to satisfy finished product demand within the due date window. From the optimal solution of BSP, we know which tank receives the product produced by the blend unit when processing PSP blend component demand order o and we can determine the very first finished product demand order o that is satisfied by the product tank. We use the end time ( Tsf j ,k ,n ) of the material flow into the tank k, start time ( Tosk ,o, n ), finish time ( Tof k ,o, n ) and a due date ( timeso - timefo ) of the first demand order o satisfied by the tank k to determine tupo as shown in Figure 6.5. Here, parameter tak is fill-draw-delay for the product tank and tupo is the maximum demurrage violation that can be incurred by blend component demand order o that is satisfied by PSP without demurrage violation of finished product due date in RBSP. Another parameter that is determined from the optimal solution of BSP is j ,i ,i . Parameter j ,i ,i is 1 if the changeover of service from task i to i' at blend unit j happens at any time during the scheduling horizon, i ,i , j , n 0 , otherwise it is 0. The n parameters Do, s , Do, s , timeso , timefo , oto,i , bo j ,o , ato , tupo , and j ,i ,i are obtained from the optimal solution of BSP and are used in formulating the production unit scheduling problem and at least MN 1 event points are needed to solve the PSP in order to obtain a feasible solution. 6.3.2. Production Unit Scheduling Problem Goal of production unit scheduling (PSP) is to satisfy demands of intermediate products ( S B ) and blend components. To this end, PSP is solved using the parameters determined from the 171 optimal solution of BSP. Formulation for PSP follows the original model expect that few new constraints are added and some existing constraints are eliminated to obtain a feasible solution that satisfy group B products demand ( rgs 0 ). The oil-refinery system under study in this work does not possess blend components storage tanks. However, to model the intermediate due dates for blend components, we have introduced Sbc dedicated storage tanks for blend component products. These artificial product tanks have zero accumulation at the end of each event-point as given by equation (4a). Thus, material balance constraint for tanks present in Chapter 4 is rewritten as equation (4a‟) for blend components tanks. sts , k , n 0, s Sbc , k K s , n N sts , k , n Kif jJ kpk s, j ,k ,n Lf k ,o , n , s Sbc , k K s , n N (4a) (4a‟) oOs In the full-scale model, the finished blend product tanks have the restriction of not loading material out of a tank when there is a material flowing into the tank. For PSP, to satisfy material balance of the blend component tanks, it is imperative to have simultaneous flow of material into and out of these tanks. Furthermore, since material flow in and out of these tanks should happen at the same time, we introduce (4b-4e) sequence constraints to align the start and finish time of flow in and out of the blend component tank. Tss j , k , n Tosk ,o , n UH 2 lk ,o , n ins , j , k , n , s Sbc , k K s , o Os , j J kpk , n N (4b) Tss j , k , n Tosk ,o , n UH 2 lk ,o , n ins , j , k , n , s Sbc , k K s , o Os , j J kpk , n N (4c) Tsf j , k , n Tof k ,o , n UH 2 lk ,o , n ins , j , k , n , s Sbc , k K s , o Os , j J kpk , n N (4d) Tsf j , k , n Tof k ,o , n UH 2 lk , o, n ins , j , k , n , s Sbc , k K s , o Os , j J kpk , n N (4e) 172 During rundown blending operations, blend unit must receive all the blend components at the same time and this requirement is captured by equations (4f-4h). Constraints (4g-4h) are included because demurrage violation variables Tlateo and Tearlyo are not be penalized in the objective function but a positive variable lateo defined in the equation (4i) is penalized. l k K p k ,o, n l k K p k , o , n , j J b , o O J , o O J , o o, bo j ,o bo j ,o , n N (4f) Tlateo Tlateo , j J b , o O J , o O J , o o, bo j ,o bo j ,o (4g) Tearlyo Tearlyo , j J b , o O J , o O J , o o, bo j ,o bo j ,o (4h) Tlateo tupo lateo , o O, ato 1 (4i) If multiple blend headers share common blend component, then at any given time, blend component can supply material to multiple blend units. However, a product tank cannot supply multiple demand orders to the same blend unit at the same time. This restriction is captured by constraint (4j). oOs , bo j ,o 0 lk ,o , n 1, k K p , s Sbc Sk , j J sc J b , n N (4j) Demand violations ( dgol 0, dgou 0 ) for blend components are not subject to penalty in the PSP objective function. However constraint (4k) ensures that the total demand for a given blend unit is satisfied for rundown blending operation at time period v. sSbc , oOs O j , k K s , ato 1,bo j ,o v Lf k ,o , n sSbc , oOs O j , ato 1,bo j ,o v Do, s , j J b , v {1, 2,..., MN } (4k) Constraint (4l) takes maximum production capacity of a blend unit into consideration by limiting total flow of blend components to the blender. 173 o O , k K p ,bo j ,o v Lf k ,o, n Rimax , j Tof k , o , n Tosk , o , n UH 1 lk , o , n , j J b , s S cj , o Os , k K s , i I j , oto ,i 1, v 1, 2,..., MN , bo j ,o v, n N (4l) To obtain a feasible solution at the end of iteration, we introduce consumption recipe constraint (4m) for blend components and finished blend-product quality constraint (4n) based on linear blending rules to maintain model linearity. sc,,min i oO , k K p , bo j ,o v Lf k ,o, n Lf k K p k ,o,n sc,,max i o O , k K p ,bo j ,o v Lf k ,o, n , j J b , s S cj , o Os , i I j , oto,i 0, v 1, 2,..., MN , bo j ,o v, n N psmin ,p oO , k K p , oto ,i 1, bo j ,o v psmax ,p Lf k ,o , n oO , k K p , oto ,i 1,bo j ,o v oO , k K p , oto ,i 1, bo j ,o v p s, p (4m) Lf k ,o , n Lf k ,o, n , j J b , i I j , s Sip , v 1, 2,..., MN , n N (4n) To restrict the solution space of the model, certain variables relating to blend components outflow from the artificial tanks are fixed to be zero as shown in equation (4o). The PSP must satisfy all the active demand orders ( ato 1 ) and at least MN 1 event points are needed to solve the PSP in order to obtain a feasible solution. We add constraints (4p-4q) in the model to facilitate longer run modes with constant blending rate in blending operations. In equation (4p), PSP is force to fulfill demand order within 3 event-points however this limit can be set higher if needed. Lf k ,o, n , lk ,o, n 0, j J b , o O j , k K p , n N , n bo j ,o (4o) lk , o , n 3, o O, ato 0 (4p) lk , o , n MN , o O, ato 0 (4q) k K p , nN k K p , nN 174 Valid inequalities for PSP are given in (4r-4v). Similar to inequality (1n), constraint (4r) reinforces that if material is flowing out of blend component tank, then it must be being produced by process units at the same event-point. Based on nature of the blend component dummy tanks, we introduce cuts (4s)-(4t) in the formulation of PSP. lk , o , n wvi , j , n , k K p , s Sbc Sk , o Os , n N iI sp , jJ i l oOs k ,o, n l jJ kpk oOs in s, j ,k ,n oOs jJ kpk k ,o, n in , s Sbc , k K s , n N s, j ,k ,n , s Sbc , k K s , n N jJ kpk (4r) (4s) (4t) Note that the demand order set pertaining to each blend header is arranged in chronologically ascending order. When bo j ,o bo j ,o ' 0 , constraint (4u) states that blend component tanks cannot satisfy demand order o before order o' and equation (4v) restates that if demand order o is satisfied at event-point n then order o' must have been satisfied at earlier even-points. lk ,o, n lk ,o ', n 1, k K p , j J b , o O, o O, bo j ,o bo j ,o , n n, n N lk ,o, n k K s , nN , n n (4u) lk ,o, n , k K p , s Sbc Sk , o Os , o Os , o o, j J b , b j ,o b j ,o , n N (4v) The objective function of the PSP is given in (4w). The second term in the objective function takes into account desire for constant blend rate and the two last terms are necessary to take into consideration the cost incurred if finished blend product demands are not met on time and the cost of storage of finished products in blend scheduling problem. The parameter tdm sS A Sbp , oOs Do, s is an upper bound on the total amount of finished blend product that can be supplied to the market. 175 z psp i , jJ i / J b , n s , k K s oO , jJ b , iI j , bo j ,o 0, oto ,i 1, k K , n ck3 sts , k , N c 4j j , n j J / J , n h c os , s , k 1 ci6,i i ,i , j , n jJ / J , iI j , i I j , i i , n m b (4w) cs8, s s , s , k , n k K m , sS k , s S k , s s , n c lateo oO , jJ b , bo j ,o 0 o O , bo j ,o 0 ck5 k , n h 7 s , s 14 o k K , n b k K m , sS k , s S k , s s ci1, j lk ,o , n ci1, j wvi , j , n o O , bo j ,o 0 c st Lf k ,o , n tdm 1 kK , oO , n For the purpose of global solution and the algorithm, we report value Z P calculated from equation (4) using optimal solution of PSP. Z PSP i , j J i , n ci1, j wvi , j , n s , k K s , n ck3 sts , k , n c i , i , j , n 6 i ,i jJ m ,iI j ,i I j ,i i , n 6.3.2.1. c 4j j , n jJ , n ck5 k , n k K , n h h c os , s, k 7 s , s k K m , sSk , s S k , s s cs8, s s , s , k , n (4) k K m , sS k , s S k , s s , n Acquiring Parameters for RBSP from the Solution of PSP To solve the restricted blend scheduling problem (RBSP), we determined certain parameters from the optimal solution of the PSP. A new set is defined as T j 1, 2,..., pn j where parameter pn j represents the total number of time periods in which blend components flow into the blend unit j from PSP and pn j lk , o , n j j v1,2,..., MN oO , bo j ,o v , k K p , nN bc . A simple algorithm is used to determine blend components flows into blend units as shown below: For j J b Initialize t 0 For n N 176 If oO , k K P ,b j ,o 0 lk , o , n 1 t t 1 For s S cj , o Os , bo j ,o 0, and k Ks If lk ,o, n 1 Uifx j , s,t Lf k ,o, n , Tsx j , s,t Tosk ,o, n , Tfx j , s,t Tof k ,o, n , tx j ,i ,t oto,i End Procedure From the optimal solution of PSP, the information regarding the blend components flow amount ( Uifx j , s ,t ) to blend unit j, start and finish time ( Tsx j , s,t , Tfx j , s,t ) of these flows, and the task that should be performed ( tx j ,i ,t ) when blend unit j is processing these blend components is obtained. Set T j and parameters Uifx j , s,t , Tsx j , s,t , Tfx j , s ,t and tx j ,i ,t are obtained from the optimal solution of PSP and will be used alongside with parameters obtained from BSP in formulating the restricted blend scheduling problem. Formula max T j can be used as a starting point in determining number j of event-points needed to obtain a feasible solution of RBSP. If blend scheduling problem can be decomposed into smaller problem than event point is determined separately for each blend scheduling sub-problem. Minimum numbers of event-points for a sub-problem are determined by finding max T j for all blend headers that belong to the sub-problem. j 6.3.3. Restricted Blend Scheduling Problem The restricted blend scheduling problem (RBSP) model incorporates all the constraints of the full-scale model and some additional constraints that are necessary to incorporate the solutions of BSP and PSP. RBSP restricts feasible solution space of BSP by fixing allowable task changeovers for multipurpose blend headers. However unlike BSP, where production in the last 177 event point is not allowed (equation (3d)), RBSP allows blending operations to happen at the last event point to provide flexibility in blending all the material supplied by upstream process units. The operating mode changeovers at multipurpose blend headers are restricted to those of BSP ( j ,i ,i ) as shown in equation (5a). 1 if j ,i ,i 0 , j, i I j , i I j , i i , n N 0 if j ,i ,i 0 i ,i , j , n (5a) We introduce a binary variable x j ,i ,t , n to capture the relationship between supply of blend components to a blend header at time period t by PSP and blending of these components at some event point n by the blend header. Variable x j ,i ,t , n is 1 if event point n allocates to time period t given tx j ,i ,t is 1. Constraint (5b) requires that each time period t must correspond to at least one event point n. Equation (5c) ensures that event point n can correspond to at most one time period t and equation (5d) forces binary variable wvi , j ,n to be 1 if binary variable x j ,i ,t , n is 1. However, if the event point n does not correspond to any time-period t then variable wvi , j ,n is forced to be zero by constraint (5e). nN , n t x j ,i ,t , n 1, j J , i I j , t T j , tx j ,i ,t 1 tT j , n t , tx j ,i ,t 1 x j ,i ,t , n 1, j J , i I j , n N x j ,i ,t ,n wvi , j ,n , j J , i I j , n N , t Tj , n t , tx j ,i ,t 1 wvi , j , n tT j , n t , tx j ,i ,t 1 x j ,i ,t , n , j J , i I j , n N (5b) (5c) (5d) (5e) 178 An assignment constraint (5f) connects fixed amount of material flowing into a blend unit j and time period t to an event point n. Continuous variable UUf j , s ,t ,n represents amount of component s supplied by PSP at time period t is consumed by a blend header j at event point n. Constraint (5g) ensures that all of the material supplied to blend header j by PSP is blended over scheduling horizon and equation (5h) determines amount of blend component s processed by blend unit j at event point n. UUf j , s ,t , n Uifx j , s ,t x j ,i ,t , n , j J b , i I j , s Sic , n N , t Tj , n t , tx j ,i ,t 1 nN , n t UUf j , s ,t , n Uifx j , s ,t , j J b , i I j , s Sic , t T j , tx j ,i ,t 1 Uif s , j , n tT j , n t , tx j ,i ,t 1 UUf j , s ,t , n , j J b , s S cj , n N (5f) (5g) (5h) iI j Equation (5i) sets binary variable x j ,i ,t , n to zero if flow of materials into blend header j at time period t is not intended for blending run-mode i, that is . tx j ,i ,t 0 . Constraint (5j) fixes amount of . material that received and blended by blend header j to zero if at given time period t, process units are not supplying blend components. x j ,i ,t ,n 0 if tx j ,i ,t 0, j b , i I j , i I j , i i , n N , t T j UUf j , s ,t , n 0 if tx iI j j ,i ,t 0, j b , s S cj , n N , t T j (5i) (5j) Constraints (5k-5l) determines the duration of each task being performed at unit j at event point n using component flow rate into blend header at corresponding time period t. Equations (5m-5n) bound the start and finish time of each task happening at event point n to the start and finish time of the corresponding time period t. 179 Uifx j , s ,t Uif s , j , n Tfx Tsx j , s ,t j , s ,t Tfi , j , n Tsi , j , n M 2 wvi , j , n x j ,i ,t , n , j J , i I j , s Sic , n N , t T j , n t , tx j ,i ,t 1 (5k) Uifx j , s ,t Uif s , j , n Tfx Tsx j , s ,t j , s ,t Tfi , j , n Tsi , j , n M 2 wvi , j , n x j ,i ,t , n , j J , i I j , s Sic , n N , t T j , n t , tx j ,i ,t 1 (5l) Tsi , j , n UH 2 wvi , j , n x j ,i ,t , n Tsx j , s ,t , j J , i I j , n N , n t , t T j , tx j ,i ,t 1 (5m) Tf i , j , n UH 2 wvi , j , n x j ,i ,t , n Tfx j , s ,t , j J , i I j , n N , t T j , n t , tx j ,i ,t 1 (5n) Constraint (5o) restricts that a time period corresponds to at most three event points similar to constraint (4p) in PSP. nN , n t x j ,i ,t , n 3, j J , i I j , t T j , tx j ,i ,t 1 (5o) To improve the performance of the model, constraint (5p) is added to restrict the feasible space of the model. Constraint (5p) results from the requirement that the scheduling problem must blend components received from PSP. RBSP must satisfy the finished blend product demands hence demand give-away variable dgol 0 and over supply of demand, dgou 0 , is penalized. Demand demurrages violations are allowed to meet minimum finished blend product demand requirement and these violation are heavily penalized. At least max T j event points are necessary to obtain a j feasible solution of the RBSP. x j ,i ,t ,n 0,UUf j , s ,t ,n 0, j J , i I j , s Sic , tx j ,i ,t 1, n N , t Tj , n t The RBSP model has the same objective function as the original full-scale model. (5p) 180 6.3.4. Integer Cuts If RBSP provides an optimal solution with due-date violations, integer cuts are generated so that they can be added to the BSP in the next iteration. These integer cuts exclude current task changeover sequence for blend headers from future blend scheduling problems. The cuts are produced by deriving alternative combinations of production task assignment to different event points such that the current production task changeover sequence i ,i, j is maintained. These alternative assignment sequences are feasible solutions of RBSP because RBSP optimizes blend scheduling problem by keeping current task changeover sequence fixed. Table 6.1 shows how alternative combinations of task allocation for blend header are obtained. For example, a BSP that has one multipurpose blend header that can perform 3 tasks to produce 3 different blend products. In the next iteration BSP is to be solved using 5 event-points and current iteration has task changeover sequence for blender (j = 1) obtained from section 6.3.1.1 using BSP solution is 3,1,1 , 2,3,1 . Possible alternative solutions of task allocation variable wvi , j , n are provided in Table 6.1 along with the original solution obtained in current iteration. The last event-point has no production taking place because we restrict solution space of BSP as mention in section 6.3.1 by fixing allocation variable wvi , j , N 0 . For the given changeover sequence, 6 alternative solutions can be possible and the total of 7 integer cuts are generated as shown in Table 6.1 Hence, multiple cuts can be generated every iteration so that the current changeover sequence is excluded from feasible solution space of BSP in the subsequent iterations of the algorithm. The cuts are only added to those sub-problems that include multi-purpose blend units. Multiple integer cuts of the form (6) are added to the BSP cut-pool for the subsequent iterations. i , j , n Q wvi , j , n tsc i , j , n NQ wvi , j , n Q tsc 1, tsc 1, 2,...TSC tsc (6) 181 Here, TSC is the total number of solutions that have been obtained to exclude task changeover sequences obtained by current and previous iterations. Set Qtsc includes assignments where wvi , j , n 1 and set NQtsc includes assignments where wvi , j , n 0 . The blend scheduling problem can be decomposed into small independent sub-problems and these independent subproblems do not share common blend components, products, and product tanks. For instances where more than one sub-problem report due-date violations, cuts are added to one of the subproblems' cut pool. It is important to note that if BSP, PSP, or RBSP are not globally optimized in steps 1, 2, and 3, adding these integer cuts to the BSP in the next iteration could potentially cut off a better global solution with lower objective value. Table 6.1 Possible task assignment solutions based on current iteration task changeovers parameter i , i , j wv 2,1, n1 Task changeover BSP solution for wv 2,1, n2 sequence: multipurpose blend unit wv 3,1, n3 Possible solutions 2,3,1 wv 1,1, n4 3,1,1 wv 2,1, n1 wv 2,1, n1 wv 2,1, n1 wv 3,1, n2 wv 3,1, n3 wv 3,1, n2 wv 1,1, n3 wv 1,1, n4 wv 1,1, n4 wv 2,1, n1 wv 2,1, n1 wv 3,1, n2 wv 3,1, n2 wv 3,1, n3 wv 1,1, n3 wv 1,1, n4 wv 1,1, n4 wv 2,1, n2 wv 3,1, n3 wv 1,1, n4 6.4. Proposed Heuristic Algorithm Detailed framework of the algorithm is as follows. The procedure for the algorithm has six steps. We define Zd in equation (7) to be the cost of due-date violations. Z Bh and Z Ph are 182 calculated using the optimal objective values of the RBSP and PSP, respectively at the hth iteration. Zd c oO 14 o Tearlyo co15Tlateo oO (7) Step 1: Solve the blend scheduling problem (BSP). Obtain parameters Do, s , Do, s , timeso , timefo , oto,i , bo j ,o , ato , tupo , and j ,i ,i .If infeasible, go to step 6. Step 2: Based on the requirement of blend components, solve the production unit scheduling problem (PSP) so that the components demand is met with minimum due date violations. The solution of PSP is Z Ph Z PSP . Obtained Set T j and parameters Uifx j , s,t , Tsx j , s,t , Tfx j , s ,t and tx j ,i ,t . Step 3: Restricted BSP (RBSP) is solved using the task changeovers sequence obtained by the solution of BSP for blend units and information about blend components flow into blend units obtained by the solution of PSP. The solution of RBSP is Z Bh Z RBSP . Step 4: The global feasible solution is Z h Z Bh Z Ph . If the optimal solution of RBSP satisfies the finished blend products demand without any intermediate due-date violations, stop the algorithm and Z Z h . If an optimal solution has demurrage violations ( Zd RBSP 0 ), go to step 5. If infeasible, go to step 6. Step 5: Develop integer cuts based on the task changeover sequence ( j ,i ,i ) of the BSP solution. Add these cuts to BSP in the next iteration. h h 1 and go to step 1. 183 Step 6: If an optimal solution of RBSP cannot be obtained without demurrage violations or BSP is infeasible, pick the best feasible solution. Z min Z h and exit the algorithm. h Initialize , Integer Cut Pool, Solve BSP Done No Feasible Yes No Yes Solve PSP Add Integer Cuts Solve RBSP , Yes emurrage No Done Done 184 Figure 6.6 Proposed heuristic algorithm 6.4.1. Termination Criterion The algorithm iterates between solving models BSP, PSP, and RBSP until a global feasible solution with minimum due-date violations and minimum total cost is obtained as presented in Figure 6.6. The optimal schedules of PSP and RBSP provide a feasible solution of the overall refinery problem. We initialized our BSP model with an empty integer cut-pool. If RBSP reports a solution that respects intermediate demand due dates, then the algorithm terminates and the current solution is kept. If RBSP provides an optimal solution with due-date violations at hth iteration, then corresponding BSP changeover solution ( ih,i, j ) is used to generate integer cuts for cut pool. RBSP can never provide a better solution than BSP. Thus, if BSP reports worse solution than the h best obtained so far by RBSP, that is Zd BSP min Zd RBSP , then the algorithm is terminated and h best global solution obtained so far is kept. BSP can report an infeasible solution when all feasible schedules from its solution space are eliminated due to the addition of integer cuts. In the case of BSP infeasibility, the algorithm terminates and the best solution from previous iterations is reported as a global solution. At the end of each iteration, the heuristic algorithm provides a feasible solution however the algorithm cannot guarantee a global optimal solution since it terminates when there is no demurrage and not when the overall objective is minimal. 6.4.2. Evaluation of scheduling horizon The evaluation of BSP, PSP and RBSP problems‟ scheduling horizon is shown in Figure 6.7. In the Figure 6.7, BSP and RBSP show schedule of a blend unit j and PSP shows schedule of materials being supplied to the blend unit ( tx j ,i ,t >0) by process units at event point n. A feasible 185 solution of PSP satisfies blend components demands requested by BSP and intermediate products demands ( S B ). However due to the production capacity limitation, PSP may not able to meet intermediate due dates of blend components unless PSP utilizes inactive demand orders ( ato 0 ). In Figure 6.7, the BSP is active only for a total of 4 event points whereas, PSP supplies material to blend unit for 5 event points. The schedule of RBSP is based upon the solution of BSP and PSP; however to meet the due dates of the finished blend products, the time period have flexibility to split. In Figure 6.7, the optimal schedule of RBSP shows that the event point n3 of PSP is split in two event points, n3 and n4 and blend header is performing task with constant blend rate for these event points. BSP PSP RBSP n1, i2 n2, i2 n1, i2 n2, i2 n1, i2 n2, i2 n4, i1 n3, i3 n3, i3 n3, i3 n4, i1 n4, i3 n5, i1 n5, i1 n6, i1 Figure 6.7 Schedule evaluation of BSP, PSP, and RBSP. Here i represents index for blend task. Both PSP and RBSP provide schedules that can be implemented without any further manipulations after algorithm is terminated. 6.5. Case Study A case study with realistic data provided by Honeywell Process Solutions (HPS) is used to illustrate the effectiveness of the proposed heuristic algorithm. The refinery produces diesel fuels, jet fuel, and other middle distillate products. The production process at Honeywell refinery consists of 2 blender headers, 13 other processing units, and 2 non-identical parallel crude distillation units (CDUs) that process two different types of crude oils. The schematic of production system is shown in Figure 4.2 and details are provided in 4.2.1. Honeywell refinery utilizes rundown blending operations at Jet blender and Diesel blender units to blend components 186 supplied directly from other process units to produce final products. Jet blender unit blends three blend component streams to produce jet fuel that can be stored in 2 product tanks. Diesel blender unit can produces three different grades of fuel: CARB diesel, EPA diesel, and Red-dye diesel by blending three different components in different proportions using three different run-modes called CARB normal, EPA normal, and Red-dye normal, respectively. There are two dedicated tanks for each grade of diesel products and one multipurpose tank that can service CARB and EPA diesel. In addition to Diesel blender, FCC is also a multipurpose unit that has two modes, max distillate mode and max gasoline mode. To apply the proposed heuristic algorithm, we decompose the case study problem into production unit scheduling and blend scheduling problem. The blend scheduling problem consists of Diesel and Jet blender units and finished product storage tanks associated with these two blenders. The production unit scheduling problem involves all other units and raw material and intermediate storage tanks. The Diesel blender does not consume the same type of blend components as the Jet blender unit and the products produced by these two blenders are serviced by different product tanks, hence, we can decompose the blend scheduling problem (BSP) into two sub-problems, the diesel blend scheduling problem (D-BSP) and the jet blend scheduling problem (J-BSP). There is a 6 hours of cleaning or maintenance downtime when the multipurpose tank service switches from lower grade of diesel product to higher grade of product. This cleaning downtime is essential to remove any sulfur contamination present in the tank before low sulfur product is sent for storage. The product tank has 4 hours of down time called fill-draw-delay for certificate of analysis preparation and to let the product settle down and mix before is shipped to the market. A minimum run-length restriction of 6 hours is imposed on all production units. The data for this case study for provided in Appendix Chapter 6. Table A6-1 provides maximum and minimum production rate for units, variable production and consumption recipe data for the production units is presented in Table A6-2, and the maximum storage capacity and what material can be stored in each tanks is given in Table A6-3. 187 6.6. Computational Results The effectiveness of the proposed heuristic algorithm in solving refinery scheduling problems is demonstrated using ten different examples for the case study presented. Each example differs in either demands, due dates, or initial hold up in the tank. All the examples are solved on a Dell® Optiplex 9020 computer with 2.90 GHz Quad Core Intel® Processor i5-4570S, CPU 2.90GHz, and 16.0 GB memory using CPLEX 12.3.0/GAMS 23.7.1. The heuristic algorithm is implemented in MATLAB R2014a and interfaces with GAMS to solve optimization problems. The maximum solution time of 8 hours and optimality tolerance of 1e-6 are used as termination criteria for all examples. The number of blend product demand orders range from 4 to 22 and with different intermediate due-dates and amount. The scheduling horizon for these problems is 240 hours (10 days). Table 6.2 provides penalty parameters used in objective function and Table A6-4 give demand data for finished blend product with intermediate due-date window. Products CARB diesel, EPA diesel, Red-dye diesel, and Jet fuel are referred to as product P1, P2, P3 and P4, respectively. Intermediate demands are bound by maximum and minimum amount and can be delivered anytime during the delivery window. Maximum delivery rate, product unloading rate to satisfy demand, is 10 (kbbl/h). Demand for distillate product for different examples is given in Table A6-5. Table 6.2 Penalty parameters in objective function Penalty Penalty Parameter Value Value Parameter 1 CCarbNormal,Diesel Blender 10 Cs7, s 40 (Unfavorable: 60) 1 CEPA Normal,Diesel Blender 7 Cs8, s 40 (unfavorable: 60) 1 CReddye Normal,DieselBlender 4 Ck9 5 1 CJetNormal,Jet Blender 5 Cs10,k 5 188 Ck2 1 (Swing tank: 3) 50 Vkmax Ck3 1 Co11 1150 Co12 800 C 4j 50 Cs13 1100 Ck5 10 (Swing tank: 40) Co14 5400 Ci6,i 100 (Unfavorable: 250) Co15 3250 Model statistics for the proposed algorithm sub-problems and full-scale problems are reported in Table 6.3. For every iteration, the minimum number of event points necessary for optimal solutions are not known prior and need to be determined by trial and error for subproblems. BSP model is solved using same number of event points during each iteration in the algorithm, as seen in Table 6.3 for examples 3 and 6. Starting point in determining the minimum number of event points for PSP is given in section 6.3.1.1 and the event points for RBSP are determined from optimal solution of PSP as mention in section 6.3.2.1. If PSP optimal solution reports due-date violations for blend component demand orders, then the number of event points necessary for obtaining optimal solution of RBSP may be higher than for corresponding PSP. Table 6.3 Model statistics Heuristic Algorithm Full Scale Model BSP model PSP model RBSP model Event pt. Event pt. Event pt. Event pt. # Int./Cont. Int./Cont. variables Int./Cont. Int./Cont. vairables Orders var. (Constraints) var. (Constraints) (Constraints) Nonzero Elem. (Constraints) Nonzero Elem. Ex. Nonzero Nonzero D-BSP Elem. J-BSP D-RBSP Elem. J-RBSP 189 1 2 5 4 2 4 5 4 360/2663 97/658 12/86 263/1599 141/878 36/382 (6387) (1698) (181) (5412) (2319) (663) 22633 6068 546 17336 8536 1798 5 4 2 4 4 5 360/2663 97/658 12/86 263/1599 110/698 47/482 (6379) (1692) (181) (5412) (1794) (844) 22569 6026 546 17344 6324 2349 5 2 5 5 5 124/822 12/86 337/2049 142/882 48/486 (2169) (181) (7381) (2327) (861) 8020 546 23843 8526 2403 5 2 5 5 4 124/822 12/86 337/2049 141/878 36/382 (2170) (181) (7381) (2311) (663) 8035 546 23853 8472 1798 5 2 5 5 4 124/822 12/86 337/2049 141/878 36/382 (2171) (181) (7381) (2309) (663) 8050 546 23853 8472 1798 5 2 5 5 5 124/822 12/86 337/2049 142/882 48/486 (2178) (181) (7381) (2326) (861) 8155 546 23843 8526 2403 5 2 5 5 4 4 4 5 360/2663 3 4 (6379) 22569 190 4 5 6 7 124/822 12/86 337/2049 141/878 36/382 (2185) (181) (7381) (2311) (663) 8260 546 23853 8472 1798 5 2 124/822 12/86 - - - (2186) (181) 8275 546 5 5 4 4 5 4 380/2771 136/886 32/202 269/1662 153/942 42/426 (6741) (2399) (481) (5819) (2525) (765) 24018 8916 1516 18733 9310 2160 6 6 4 6 9 7 484/3453 181/1138 38/238 426/2610 322/1858 98/850 (8644) (3227) (611) (10364) (5464) (1686) 31765 12507 1978 33812 24234 5369 7 4 7 8 8 226/1386 44/274 518/3120 299/1722 131/1070 8 (4038) (761) (13029) (5011) (2284) 680/4737 16326 2524 42828 21059 7794 (12260) 7 4 7 9 8 47983 226/1386 44/274 518/3120 341/1946 131/1070 (4087) (761) (13029) (5750) (2284) 17355 2524 42842 25360 7794 7 7 4 7 8 8 640/4383 274/1622 44/274 509/3120 355/1990 131/1118 6 8 10 13 191 8 9 10 (11628) (5057) (761) (13349) (6116) (2284) 44759 20588 2524 43662 25900 7842 7 6 4 6 7 7 682/4591 256/1522 50/310 417/2436 332/1870 120/1046 (12669) (4898) (931) (8739) (5916) (2219) 49601 19356 3154 28529 24332 7739 8 7 6 7 8 9 841/5509 328/1886 100/558 518/3120 418/2290 199/1548 (16222) (6512) (2061) (13029) (7722) (3929) 67119 27371 7790 42856 33814 15617 10 7 8 7 9 10 1117/7125 340/1946 166/870 512/3021 493/2658 261/1914 (22251) (6801) (3823) (12027) (9183) (5672) 100257 28643 16250 39592 42180 24780 16 19 22 Table 6.4 Computational results Objective ( Z ) Ex. Full Scale Heuristic Algorithm PSP D-BSP J-BSP Total PSP D-RBSP J-RBSP Total 1 109.67 276.21 64.90 450.78 110.88 276.22 60.74 447.84 2 105.16 412.48 81.00 598.64 129.57 421.55 101.46 652.58 3 116.28 14530.12 57.59 14704.00 119.49 15751.44 58.28 15929.21 4 117.66 416.50 58.65 592.81 115.20 427.25 61.11 603.56 5 144.74 294.00 65.46 503.19 150.53 318.00 81.84 550.37 192 6 154.27 692.85 69.52 916.64 155.67 576.43 80.82 812.92 7 150.54 303.00 59.07 512.61 149.89 309.36 69.26 528.51 8 140.37 446.00 51.00 637.37 156.40 461.15 58.21 675.76 9 165.39 562.46 67.50 795.34 162.48 567.92 71.58 801.99 10 150.69 868.78 79.17 1098.64 166.92 462.31 74.62 703.85 The objective of the proposed algorithm is to obtain a global feasible solution that satisfies minimum demand requirements while minimizing product shipment commitments violations, inventory costs, product giveaway, product tank heel quantity, and undesired mode changeovers at multipurpose blend headers and product tanks. Table 6.4 gives the objective function of full-scale model and heuristic approach. The minimum demand violations are allowed and not fixed to zero in full-scale problem in order to find the optimal solution. To compare quality solution obtained by integrated approach versus decomposition approach, the optimal solution of the integrated problem is used to calculated objective value for each sub-problems. The computational performance results of the full-scale model and the proposed algorithm are shown in Table 6.6. The proposed algorithm is able to obtain a solution at significantly less computational expense than the full-scale model and in some instances full scale model failed to find optimal solutions after 8 hours. For the largest examples with 22 intermediate demand orders, the full-scale has optimality gap of 82.82% after 8 hours whereas the heuristic approach is able to obtain a solution after 865 seconds. For examples 3 and 6, heuristic approach takes more than one iterations before the algorithm terminates and the progress of the algorithm is shown in detail in Table 6.7. Majority of computational effort in the algorithm is spent solving for BSP while RBSP consume less computational time compared to both BSP and PSP. In our work, blend scheduling sub-problems are not solved in parallel however to reduce computational time, they can be solved in parallel. Of two blend scheduling sub-problem, Diesel blend scheduling 193 sub-problem is more complicated due to presence of multipurpose blend header, hence is much harder to solve than Jet blend scheduling sub-problem. The objective function results obtained using decentralized approach are comparable to that of centralized approach for all examples as shown in Table 6.4. However, for example 1, an optimal solution of integrated full-scale model is higher than that of heuristic approach solution. This happens because our objective function includes a term i , j J i , n ci1, j wvi , j , n . Thus, the objective value of sub-problems is highly dependent on the number of event points used to solve scheduling model. This term is included in objective function as to minimize number of different run modes hence to maintain constant blend rate for longer length of time. In example 1, the fullscale model is solved using 5 event points and diesel and jet blenders are active for all 5 event points, where as in heuristic approach, only diesel blender is active for 5 event points and jet blender for 4 event points since only D-RBSP is solved using 5 points. zcost k K p , o , n ck2 lk , o , n s , k K s ci6,i i ,i , j , n jJ , iI j , i I j ,i i , n m c sS f c 4j j , n 9 k K m ck5 k , n jJ h , n k K h , n cs7, s os , s , k k K , sS k , s S k , s s m c k K m , sSk , s S k , s s , n 13 s ck3 sts , k , N std s , s , k , n k K m , sS k , n cs8, s s , s , k , n k K , sS k , s S k , s s , n m c mhs , k , n co11dg ol co12 dg ou 10 s,k oO (8) oO rg s c Tearlyo c Tlateo oO 14 o oO 15 o To effectively compare the quality of the solution between full-scale and heuristic approach, we calculate value Zcost given equation (8) using optimal solution. Here, Zcost is almost same as optimization problem objective function but the term i , j J i , n ci1, j wvi , j , n is not considered. When Zcost value is compared for heuristic and full-scale approach, as expected, the fullscale model optimal solution gives lower cost ( Zcost ) than the proposed algorithm as seen in Table 6.5. Proposed algorithm is successful in obtaining a feasible solution that satisfies minimum and 194 maximum demand requirement while meeting intermediate due-dates for every example for which such a feasible solution exists. Furthermore, heuristic approach is able to provide solution without product giveaway or demand violations for all examples except example 3. For example 3, the full-scale problem reports an optimal solution is with due-date violations and product giveaway. Product giveaway occurs when premium quality product must be given away for the regular product price to meet the regular product demands shipment commitment. The largest example with 22 intermediate demand orders, heuristic provides feasible solution that satisfy demand and intermediate due-dates while minimizing unfavorable blend mode changeovers and product giveaway whereas the full-scale model provides feasible solution with product giveaway and higher unfavorable blend mode changeovers at multipurpose blender. Table 6.5 Computational results Objective ( Zcost ) E Full Scale Heuristic Algorithm x PSP D-BSP J-BSP Total PSP D-RBSP J-RBSP Total 1 109.67 247.21 39.90 396.78 110.88 247.22 40.74 398.85 2 105.16 384.48 61.00 550.64 129.57 390.55 76.46 596.58 3 116.28 14502.00 37.59 14656.00 119.49 15723.44 38.28 15881.22 4 117.66 388.50 38.65 544.81 115.20 399.25 41.11 555.56 5 144.74 256.00 40.46 440.19 150.23 256.00 46.84 453.37 6 154.27 647.85 39.52 841.64 155.67 519.43 40.82 715.91 7 150.54 261.00 29.07 440.61 149.89 263.36 29.26 442.51 8 140.37 411.00 26.00 577.37 156.40 416.15 28.21 600.76 9 165.39 504.46 32.50 702.34 162.48 509.92 31.58 703.99 10 150.69 809.78 39.17 999.64 166.92 414.31 34.62 615.85 195 Except for examples 3 and 6, the heuristic algorithm is able to obtain a feasible solution without demurrage in first iteration. The algorithm takes 2 iterations to obtain a global feasible without demurrage in case of the example 6. Even for example 6, heuristic algorithm is able to find a solution in about 564 seconds while full-scale model is not even able to find an optimal solution after 8 hours. Table 6.6 Computational performance results Heuristic Algorithm Full Scale Model CPU time (seconds) Ex. CPU Gap time (s) (%) 1 108.59 2 249.85 3 1939.09 D-BSP J-BSP PSP D-RBSP J-RBSP Total 0.00 0.44 0.06 2.53 0.20 0.06 3.29 0.00 0.50 1.01 2.76 0.08 0.20 4.56 19.11 0.047 5.32 0.125 0.046 12.25 0.047 6.24 0.327 0.031 12.86 0.063 6.55 0.250 0.062 0.00 117.09 15.38 0.062 13.10 0.203 0.047 8.33 0.032 4.87 0.343 0.062 11.25 0.093 - - - 4 457.27 0.00 1.23 0.047 5.73 0.16 0.031 7.192 5 14716.62 0.00 2.26 0.11 18.31 11.89 0.39 32.96 246.56 0.047 90.71 14.77 1.72 6 28800 77.98 141.63 0.047 27.32 40.42 1.23 564.46 7 28800 9.05 56.52 0.063 28.64 1.37 0.86 87.46 8 28800 69.92 16.51 0.078 33.06 14.62 2.34 66.44 9 28800 74.71 296.73 0.250 37.25 88.41 8.00 430.64 196 10 28800 82.82 232.10 1.030 51.56 444.87 135.85 865.40 Evolution of the algorithm for example 6 is shown in Table 6.7. In the first iteration, Diesel RBSP obtains optimal solution with intermediate due-date violations. Since only Diesel blend scheduling sub-problem has multi-purpose blender, we produce integer cuts for next iteration using the current iteration task changeover sequence of diesel blender. Diesel blender can perform three tasks (i1, i2, and i3) to produce three different grades of diesel blend products. Changeover sequence ( i ,i, j ) in first iteration is (i3, i1), (i1, i2) and (i2, i3) and Diesel-BSP sub problem in next iteration is solved using 7 event-points. Total of 49 alternative solutions are obtain from the changeover sequence as described in section 6.3.4 and thus 49 integer cuts are added to D-BSP in 2nd iteration. The algorithm is terminated after 2nd iteration when a global feasible solution without demurrage costs is obtained. Table 6.7 Algorithm progress and termination Ex. Iter. Objective( Z ) Integer Demurrage Cost ( Zd ) Cuts D-BSP J-BSP PSP D-RBSP J-RBSP Total 2120.64 16.00 107.01 16497.52 63.97 16668.50 886.36 0 0 15247.44 0 15247.44 2120.64 16.00 119.51 16481.71 55.83 16657.05 886 0 0 15247.44 0 15247.44 5467.63 16.00 119.49 15751.44 58.28 15929.21 4963.64 0 0 15247.44 0 15247.44 5467.64 16.00 107.00 15767.25 62.86 15937.11 4963.64 0 0 15247.44 0 15247.44 1 Generated 1 2 1 3 3 7 4 7 197 8207.00 16.00 119.50 20440.81 55.83 20616.14 7150.00 0 0 19383.81 0 19383.81 24006.24 16.00 - - - - 22750.00 0 518.00 39.00 154.89 24852.47 79.81 25087.17 0 0 0 24248.63 0 24248.63 518.00 39.00 155.67 576.43 80.82 812.92 0 0 0 0 0 0 5 1 6 1 - 49 6 2 - The heuristic algorithm progress for example 3 is reported in Table 6.7 and the algorithm terminates during 6th iteration and the best solution found so far is kept. For example 3, in the first iteration, D-RBSP sub-problem reports a solution with due date violations thus integer cuts are produced to eliminate the current changeover sequence and are added onto the D-BSP integer cut-pool. In 6th iteration the objective of D-BSP is 24006 and demurrage cost is 22750 which is worse than the best objective of 15929 and demurrage 15247 obtained in 3rd iteration. Thus, the algorithm terminates during 6th iteration because the current D-BSP optimal solution is worse than the best D-RBSP solution obtained so far and RBSP can never do better than its corresponding BSP. The solution obtained by 3rd iteration is reported as the final solution since it has the lowest overall cost of all 5 iterations and lowest demurrage cost. Total of 17 integer cuts are generated at the end of 5th iterations and it takes only about 120 seconds for algorithm to terminate compared to 1939 seconds for full-scale model. For example 3, the full-scale integrated optimal solution reports demurrage violation for a blend product demand order to be 4.32 hours whereas the proposed heuristic approach solution gives violation to be 4.69 hours. Furthermore, the solution obtained using the heuristic algorithm has the same product giveaway cost as fullscale optimal solution. 198 The performance of our heuristic algorithm is significantly better than that of full-scale model and decomposed problems are able to produce schedules that are similar to that of full-scale model. Figure 6.8 and Figure 6.9 shows Gantt chart for production schedule for example 5 obtain using integrated full-scale model and heuristic model, respectively. The decomposition approach provides similar schedule as integrated approach. Since the refinery under study has blend headers that receive blend components straight from upstream process units, blend headers must always be active when process units are sending materials. This continuous process characteristic of rundown blending is clearly observed in the Gantt charts. It is preferable for blend headers to blend material at constant blend rate. In our heuristic approach, constant blend rate is achieved when material being sent by upstream processes at given time-period is processed during different event-points when restricted blend scheduling problem is solved using production unit scheduling problem results. In Figure 6.9, for the last two event points when Jet blender is using constant blend rate of 0.4826 when blending 106.8 kbbl and 44.5 kbbl amount of jet fuel. Similarly, even though diesel blender has two smaller run modes, blend rate is kept constant between 73.98 to 99.21 hour, and 99.21 to 132.45 hour. 199 Figure 6.8 Gantt chart for Example 5 obtained using integrated full-scale model 200 Figure 6.9 Gantt chart for Example 5 obtained from the heuristic algorithm The work in this chapter is based upon the linear blending rules for finished products and this approach does not reflect the non-linearity of the mixing rules. When the non-linear mixing rules are considered, the scheduling model becomes MINLP instead of MILP which is much harder in terms of computational complexity. The approach proposed here can be utilized as a 201 preprocessing step before tackling complicated MINLP problem. To reduce the complexity of the MINLP scheduling problem, solution space of the non-linear model can be restricted based upon the solution of the pre-processing step. In pre-processing step, the corresponding MILP model of MINLP will be first solved using the heuristic algorithm presented in this work. The obtained feasible solution will provide information on the blend mode changeover sequences for multipurpose blend headers necessary for meeting minimum demand requirements and intermediate due-dates. By using this information, MINLP feasible solution can be restricted to allow tasks changeovers that are only present in the corresponding MILP. To further reduce complexity of MINLP, other binary and allocation variables can be fixed to the solution obtained from the heuristic algorithm. Proposed algorithm can also be extended to refineries that have blend component storage tanks by making minor changes in PSP model. 6.7. Summary In this chapter, we present valid inequalities for the large-scale oil-refinery scheduling problems that improve the computational expense necessary to obtain optimal solution. An integrated refinery problem is decomposed into production unit scheduling problem and finished product blending and delivery problem via splitting blend components streams. A heuristic algorithm is proposed based on the decomposed network to obtain a feasible solution of the original problem that satisfies minimum demand requirements while minimizing due-date violations and maximizing performance. It is demonstrated on realistic diesel refinery case study that the proposed algorithm provides feasible solutions that are closer to the optimal solution with significantly less computational effort than that required by the full-scale model. Nomenclature Indices 202 i Tasks j Production units k Storage tanks n Event points o Product order p Properties s States Periods relating to blend components flow into a unit (calculated from t optimal solution of BSP) Periods relating to blend components flow into a unit ( calculated from v optimal solution of PSP) Sets Ij Tasks which can be performed in unit j I sc / I sp Tasks which can consume/produce material s J Production units Jb Blend units J sc / J sp Units that consume/produce material s Ji Units which are suitable for performing task i Units that can produce all the same products as some other unit in the Jh refinery Jm Units which are suitable for performing multiple tasks J kkp / J kpk Units that consume/produce material s stored in tank k J seq j Units that follow unit j (no storage in between) K Storage tanks 203 Kh tanks that can store the same products as some other tank in the refinery pk K kp j / Kj Tanks that store material consumed/produce by unit j Km Multipurpose tanks that can store multiple materials Kbp Tanks that can store finished blend products Kp Tanks that can store products Ks Tanks that can store material s N Event point within the time horizon O Orders for products that are stored in tanks P Product Properties S States SA Group A products, stored in tanks SB Group B products, not stored in tanks Sbp finished blend products, stored in tanks Sk Materials that can be stored in tank k Sbc Blend components Sbcj Blend components that can be consumed by blend header j Sic / Sip Materials that can be consumed/produced by task i S cj / S jp Materials that can be consumed/produced by unit j Time periods within RBSP during which blend unit j receives component Vj flow Parameters ato 1 if the order o of PSP has to be satisfied by PSP. bo j ,o Order o of PSP is needed by blend unit j 204 Do, s , Do, s Demand limit requirement for order o and product s that is stored in tank max_rates Maximum production rate of material s nj Number of total event points used by blend unit j in BSP model solution oto,i 1 if task i is performed when order o of PSP is processed by BSP. Ps, p Property p of blend component s max Psmin , p / Ps , p Minimum/maximum specification of property p of blend product s Total number of time period in which blend components flow into the blend pn j unit j from PSP rs Demand of the final product s at the end of the time horizon max Rimin , j / Ri , j Minimum/ maximum rate of material be processed by task i in unit j RLi Minimum run length for task i RU kmin / RU kmax Minimum/maximum rate of product unloading at tank k stos,k Amount of state s that is present at the beginning of the time horizon in k tak Fill draw delay for product tank k Start and finish time of blend component s flow to blend unit j at time Tsx j , s,t / Tfx j , s,t period t tupo Allowable due-date violations of order o without any penalty in PSP tx j ,i ,v At time period v, blend unit j is performing task i UH Available time horizon Amount of blend component s received by blend unit j from PSP at time Uifx j , s ,v period v Vkmax Maximum available storage capacity of storage tank k Vkheel Maximum heel available for storage tank k 205 yos , k 1 if the material s is present at the beginning of the time horizon in k j ,i ,i 1 if changeover of task happen from i to i’ at blend unit j in BSP max smin ,i / s ,i Proportion of state s produced/consumed by task i, Variables Binary Variables wvi , j , n Assignment of task i in unit j at event point n ins , j ,k ,n Assigns the material flow of s into storage tank k from unit j at point n Assigns the starting of product flow out of product tank k to satisfy order o lk ,o,n at event point n outs , k , j , n Assigns the material flow of s out of storage tank k into unit j at point n Denotes that process material received by blend unit j at period v is x j ,i , v , n processed by task i at event point n ys , k , n Denotes that material s is stored in tank k at event point n Positive variables bps ,i , j , n Amount of material s produced task i in unit j at event point n bcs ,i , j , n Amount of material s undertaking task i in unit j at event point n dgol Minimum demand quantity give-away term for order o dgou Maximum demand quantity give-away term for order o H Total time horizon used for production tasks JJf s , j , j , n Flow of state s from unit j to consecutive unit j’ for consumption at point n Kif s , j ,k ,n Flow of material s from unit j to storage tank k event point n Kof s , k , j , n Flow of material s from storage tank k to unit j at point n lateo In PSP, due date violation of blend component demand order o 206 Lfo, k , n Flow of final product for order o from storage tank k at event point n mhs , k , n Maximum heel give-away term for product tanks rgs Minimum demand quantity give-away term for Group B product s Rif s , k , n Flow of raw material to storage tank k event point n sts , k ,n Amount of state s present in storage tank k at event point n Amount of state s that is downgraded to state s‟ in storage tank k at event std s , s, k , n point n Tearlyo Early fulfillment of order o than required Tfi , j ,n Time that task i finishes in unit j at event point n Tlateo Late fulfillment of order o than required Tosk ,o, n Time that material starts to flow from tank k for order o at event point n Tof k ,o, n Time that material finishes to flow from tank k for order o at event point n Tsi , j ,n Time that task i starts in unit j at event point n Tsf j , k ,n Time that material finishes to flow from unit j to tank k at event point n Tsf k , j , n Time that material finishes to flow from tank k to unit j at event point n Tss j , k , n Time that material starts to flow from unit j to storage tank k Tssk , j , n Time that material starts to flow from tank k to unit j at event point n Uif s , j ,n Flow of raw material s to production unit j at point n Uof s , j , n Flow of product material s from unit j at point n UUf j , s ,t ,n Amount of state s received by unit j at period v is processed at event n j ,n For unit j, 1 if the unit becomes active for very first time at event point n k , n For tank k, 1 if the tank becomes active for very first time at event point n 207 s , s , k , n Continuous 0-1 variable, 1 if material in tank k switchover service from s at event point n to s‟ at later event point os , s, k Continuous 0-1 variable, 1 if material in tank k switchover service from s to s‟ i , i , j , n Continuous 0-1 variable, 1 if task at unit j changes from i at event point n to i’at later event point. 208 Chapter 7 7. Augmented Lagrangian Relaxation Approach for Refinery Scheduling To improve quality of the decision making in the refinery operations, coordination between frontend of a refinery to end-products blending is necessary. Finished product blending and delivery operations and production unit operations present unique challenges each and lead to a complex scheduling model. In this work, we present a unified framework for decomposition via augmented Lagrangian relaxation. The coupling constraints of two operations system are relaxed using augmented Lagrangian relaxation technique and the full-space model is decomposed into production unit scheduling relaxed sub-problem and finished product blending and delivery scheduling relaxed sub-problem. The non-separability terms introduced by augmented penalty are resolved using diagonal quadratic approximation method. A general decomposition algorithm that addresses real-world features such as parallel blend headers, multipurpose product tanks, product giveaway, and blending operations with/without components storage. Applicability of proposed approach is presented on by addressing four different refinery network configurations: no blend component tanks, parallel blend units and no component tanks, component tanks present and no direct flow to blend unit, and direct flow to blend unit and components storage options. 7.1. Introduction Lagrangian decomposition (LD) method developed in previous chapter addresses scheduling of refinery operations with rundown blending (no blend component tanks) and proposed restricted LD is effective addressing large scale problems. In this chapter, a unified decomposition strategy based on Augmented Lagrangian relaxation method is developed that can address different refinery configurations effectively. In ordinary Lagrangian decomposition, duality gap exists between solution of dual problem and the solution original problem in presence of integer variables or other non-convexities. Thus, heuristic procedures are needed to obtain feasible 209 solution at the end of each iteration. To overcome this problem, an augmented Lagrangian relaxation (ALR) method is adopted. ALR method has been used in several applications (Andreani et al., 2008; Fortin & Glowinski, 1983; Gupta et al., 2001; Z. Li & Ierapetritou, 2010a; Nishi et al., 2008; Tosserams et al., 2006; Tosserams et al., 2008) and strong convergence property of ALR algorithm is proved by (Andreani et al., 2008). One of the drawbacks of ALR method is introduction quadratic term by the augmented penalty term. This nonseparability can be resolved using Diagonal Quadratic Approximation (DQA) method (Ruszczynski, 1995), the Block Coordinate Decent (BCD) method (Bertsekas, 1995; Bertsimas & Sim, 2003), and the Alternating Direction Method (Bertsekas & Tsitsiklis, 1989). In this chapter, we apply augmented Lagrangian optimization algorithm to decompose refinery operations scheduling problem into production unit scheduling sub-problem and finished product blending and delivery scheduling sub-problem. This chapter is organized as follows. Section 7.2 presents solution strategy, section 7.3 proposes detailed Augmented Lagrangian decomposition (LD) algorithm, and strengthening of decomposition algorithm is presented in section 7.4. Detailed algorithm steps for augmented Lagrangian relaxation are given in section 7.5. Computational results for several different refinery configurations on which algorithm was applied are given in section 7.6 and the chapter concludes with summary in section 7.7. 7.2. Solution Strategy 7.2.1. Augmented Lagrangian method Augmented Lagrangian relaxation is an appropriate decomposition approach when the problem under study has complicating constraints that upon relaxation leads to smaller subproblems that are easier to solve. This approach is suitable for integrating two independent scheduling models without needing to obtain full-scale model or using heuristics to find feasible solution. We briefly present a general augmented Lagrangian optimization algorithm for a following linear problem: 210 Initial Problem (IP) z Min cx dy s.t. Hx b 0 Ax By 0 x n , y q Where, c n , d q , b m , A and B are m n and m q matrices, respectively. To decompose the model (IP) into sub-problems, we introduce duplicating variables xx and coupling constraints x xx 0 . z Min cx dy s.t. Hxx b 0 Ax By 0 x xx 0 x n , xx n , y q When coupling constraint are relaxed and incorporated into the objective function. In addition to Lagrangian term, augmented penalty terms are introduced to reduced duality gap. Relaxed Problem (RP) zR Min cx dy T x xx x xx 2 s.t. Hxx b 0 Ax By 0 x n , xx n , y q Where, represents Lagrangian multipliers and is a positive penalty parameters. The augmented penalty term cannot be decomposed as it contains an inseparable cross penalty terms ( x xx ). To resolve the non-separability issue, diagonal quadratic approximation (DQA) method is used to linearize the cross-product quadratic term around the tentative solution. (Y. Li et al., 2008) ( x , xx ). x xx x xx x xx x xx 2 2 2 2 Thus with above substitution for the non-separable term, the relaxed problem (RP) can be decomposed into sub-problem (RP1) and sub-problem (RP2). 211 Subproblem Problem (RP2) Subproblem Problem (RP1) Z R1 Min cx dy T x x xx 2 s.t. Ax By 0 x , y n zP 2 Min T xx x xx x xx 2 2 s.t. Hxx b 0 q xx n Resulting decomposed models (RP1) and (RP2) are quadratic optimization problems due to quadratic terms in the objective function. The procedure for augmented Lagrangian optimization is as follows: Step1. Initiate the Lagrangian multiplier (0) and penalty parameter (0) 0 , m 0 Step 2. For given ( m) , ( m) , xx , solve sub-problem (RP1). Update x x . Step 3. Solve sub-problem (RP2) based on ( m) , ( m) and x . Update xx xx . Step 4. Update the Lagrangian multipliers using multiplier method ( m 1) ( m ) m x xx . m Step 5. If g m g m 1 , update m1 m . Here, (0,1) . For convex objective function, 1 is strictly necessary and for fast convergence, 2 3 range is recommended. (Tosserams et al., 2006) Step 6. If g , the feasible solution is calculated. Here, 0 Otherwise, m m 1 . Step 7. If iteration number m is larger than the pre-specified number, terminate. Otherwise, go to step 2. Here, g x xx , is a coupling constraint violation matrix. Penalty parameter ( ) is updated when the constraint violation is not decreased by a factor . Algorithm is terminated when the tolerance is met for constraints violations and feasible solution is found or predefined maximum iteration limit is reached. The augmented Lagrangian optimization algorithm always converges to a feasible solution for . 7.2.2. Decomposition Strategy Large scale MILPs such as problem (P) proposed in Chapter 4, require specialized solution algorithms that can take into account unique characteristics of unit specific event – points 212 in presence of multi-purpose blend units, and intermediate due-dates. We apply augmented Lagrangian optimization algorithm for solving model (P) and propose modified sub-problems that helps algorithm to converge to an optimal or near optimal solutions. To obtain sub-problems, refinery network is decoupled following the concept of spatial decomposition (e.g. (Karuppiah et al., 2008)). The network is split into two sub-networks, one belonging to production unit operations and other belonging to finished product blending and delivery operations. Production units PN BN kCf sP,t kCf sB,t kCTf sP,t , kCTssP,t kCTf sB,t , kCTssB,t Cf sP, j ,t Cf sB, j ,t CTf jP,t , CTs Pj ,t CTf jB,t , CTs Bj ,t Production Unit Operations Component tanks Blenders Blending and delivery Operations Figure 7.1 Spatial decomposition of a refinery network To picture physical split, imagine cutting pipelines transferring blend components from production units to blend units or blend component tanks. The variables corresponding to the split pipelines correspond to complicating variables in full-scale model and these variables are amount of material flow and start and finish time of the flow between production operations sub-structure (PN) and blending operations sub-structure (BN). In original model (P), flow of blend component s from production unit j to blend unit j at event point n ( JJf s , j , j , n ), flow of component s from production unit j to blend component storage tank k at event point n ( Kif s , j ,k ,n , ins , j , k , n ), and start and finish timing variables for flow of material between unit and component tanks ( Tss j , k , n , Tsf j , k , n ), are complicating variables. Other variables in model (P) are called non-linking 213 variables since they are separate for both sub-structures PN and BN. Using the approach presented in section 7.2.1, the complicating variables are duplicated, coupling constraints are introduced to model (P) and then these coupling constraints are relaxed in Augmented Lagrangian fashioned using diagonal quadratic approximation method to eliminate non-separable terms. This allows us to decompose the relaxed model (RP) shown below into two models corresponding to sub-networks PN and BN. min L( , , ) z sSbc , j , j , n s , j , j , n JJf sB, j , j , n JJf sP, j , j , n j , k , n Tss j , k Kbc , n B j , k , n Tss P j , k , n j , k K bc , n JJf Kif P JJf sB, j , j , n JJf s , j , j , n sSbc , j , j , n P Kif sB, j , k , n Kif s , j , k , n sSbc , j , k , n sSbc , j , k , n s , j , k , n Kif sB, j , k , n Kif sP, j ,k ,n j , k , n Tsf jB, k , n Tsf jP, k , n 2 B s , j , j , n 2 B s , j , k , n JJf sP, j , j , n JJf Kif sP, j , k , n 2 B Kif 2 P s , j , j , n JJf s , j , j , n B P s , j , k , n Kif s , j ,k ,n Tss B j , k , n Tss Pj , k , n Tss B j , k , n Tss j , k , n Tsf B Tsf jP, k , n Tsf B Tsf P Tss Bj , k , n Tss j , k , n j , kKbc , n P Tsf jB, k , n Tsf j , k , n j , kKbc , n 2 2 j , k , n 2 2 j , k , n s.t. constraints corresponding and variables in (P) proposed in Chapter 4 P 2 2 P j , k , n 2 2 (RP) However, to avoid introducing large set of complicating variables ( JJf s , j , j ,n , Kif s , j ,k ,n , ins , j , k , n , Tss j , k , n , Tsf j , k , n ), we aggregate the blend component s streams from multiple production units supplying BN into one stream supplying BN as seen in Figure 7.1. Similarly, multiple inlet streams for a blend component s that has storage option available, is aggregated into one inlet stream which is then split into multiple stream connecting different component tanks (Figure 7.1). New the complicating variables are reduced to ( Cf s, j ,n , KCf s ,n , KCTss , n , KCTf s , n ). Here, Cf s, j ,n is flow of component s from sub-structure PN into blend unit j in BN an event-point n . For blend component that can be stored in component tanks, KCfn s , n is flow of component s from PN to BN at event-point n , and KCTss , n / KCTf s , n is start/finish time of flow of component s from PN to BN at event-point n . 214 Complicating variables are in terms of continuous time representation and notion of unit specific event points. When augmented Lagrangian algorithm is implemented, this types of model leads to sub-optimal solutions and oscillations in objective function in the beginning of the algorithm, especially for a refinery configuration that has no blend component storage options and multipurpose blend units. To eliminate this behavior, we introduce a new transfer event-point t that is only active if there is flow of material between production unit operations and blending operations and the complicating variables can we replaced with ( Cf s , j ,t , KCf s ,t , KCTss ,t , KCTf s ,t ). The detail information regarding new transfer event-point t and its relationship with event-point n is provided in detail in section 7.3. Furthermore, the start and finish time of component s flow between PN and a blend unit j are not defined explicitly in full-scale model and therefore these complicating variables ( CTs j ,t , CTf j ,t ) need to be defined in decomposed models. The timing variables for a component s flow between PN and a blend unit j at transfer event- point t is defined as CTs j ,t and CTf j ,t instead of CTss , j ,t and CTf s , j ,t because all components that are supplied directly to the blend unit must have same start and finish times due to continuous production. Using the updated set of complicating variables ( Cf s , j ,t , KCf s ,t , CTs j ,t , CTf j ,t , KCTss ,t , KCTf s ,t ), we obtain a new relaxed problem (LRP) where old complicating variables ( JJf s , j , j , n , Kif s , j ,k ,n , ins , j , k , n , Tss j , k , n , Tsf j , k , n ) connecting two sub-structures and related constraints are eliminated from model (RP). 215 min f z sSbc , jJ BU , tT s , j ,t Cf sB, j ,t Cf sP, j ,t KCf s ,t sSbc , t CTs j ,t CTs Pj ,t j ,t CTf jB,t CTf jP,t B j ,t jJ BU ,t KCf sP,t s ,t KCTssB,t KCTssP,t s ,t KCTf sB,t KCTf sP,t B s ,t Cf Cf Cf Cf CTs CTs CTs CTs CTs CTs CTf CTf CTf CTf CTf CTf KCf KCf KCf KCf KCf KCf KCTs KCTs KCTs KCTs KCTs KCTs KCTf KCTf KCTf KCTf KCTf KCTf P B Cf s , j ,t Cf s , j ,t sSbc , jJ BU , t sSbc , t 2 2 P j ,t j ,t 2 2 B s ,t P s ,t s ,t B s ,t P s ,t 2 P 2 B s ,t s ,t B P s , j ,t s , j ,t B P j ,t j ,t B P s ,t s ,t 2 B s ,t P s ,t B P s ,t 2 2 P 2 s ,t B s ,t 2 s ,t 2 P 2 j ,t B j ,t P j ,t B j ,t P B s ,t P s , j ,t B j ,t 2 P B j ,t 2 B s , j ,t P 2 j ,t B j ,t jJ BU , t 2 2 B P s ,t s ,t (LRP) s.t. constraints corresponding to non-linking variables in (P) Following the augmented Lagrangian decomposition approach presented in section 7.2.1, model (LRP) can be decomposed into two relaxed sub-problems (LBSP) and (LPSP) such that model (LBSP) includes equations and variables pertaining to sub-network BN, while model (LPSP) includes equations and variables corresponding to the structure PN. The bounds of nonlinking variables in both the both sub-problems are same as original full-space model (P) whereas the original linking variables are eliminated from the original full-space problem (P) and will be replaced with new variables that will be defined in section 7.3. The two models (LBSP) and (LPSP) are as follows: min f BSP z sSbc , jJ s , j ,t Cf sB, j ,t BU j J ,t sSbc , t s ,t BU CTs Bj ,t j ,t CTf jB,t ,t B s ,t P Cf sB, j ,t Cf s , j ,t P B KCf s ,t KCf s ,t sSbc , t j ,t KCf s ,t KCTssB,t s ,t KCTf sB,t sSbc , jJ BU , t 2 P B CTs j ,t CTs j ,t jJ BU , t KCTs 2 B s ,t P KCTs s ,t 2 CTf jB,t CTf KCTf 2 s.t. constraints corresponding to units and connections in BN B s ,t P KCTf s ,t P j ,t 2 2 (LBSP) 216 min f PSP z s , j ,t sSbc , jJ BU , t sSbc , t s ,t Cf sP, j ,t jJ BU , t j ,t CTs Pj ,t j ,t CTf jP,t KCf s ,t KCTssP,t s ,t KCTf sP,t P s ,t Cf Cf CTs CTs CTs CTs CTf CTf CTf CTf Kif KCf KCf Kif KCTs KCTs KCTs KCTs KCTf KCTf KCTf KCTf B P Cf s , j ,t Cf s , j ,t BU sSbc , jJ , t 2 B j ,t 2 P j ,t j ,t 2 B P s ,t s ,t B s ,t P s ,t B P s ,t sSbc , t s ,t B P s , j ,t s , j ,t 2 P 2 j ,t B j ,t P j ,t B jJ BU , t 2 B P j ,t j ,t 2 B P s , j , k , n s , j , k , n 2 2 B s ,t P 2 s ,t 2 B P s ,t s ,t 2 s.t. constraints corresponding to units and connections in PN (LPSP) Here, s, j ,t , s, j ,t , s, j ,t , s,t , s,t , and s,t represent Lagrangian multipliers and is the corresponds to augmented Lagrangian penalty term. The scheduling model (LBSP) and (LPSP) min min can be solved using different number of event points, epPN and epBN corresponds to the minimum number of event-points ( n ) required to obtain optimal solution of (LPSP) and (LBSP), min respectively. The cardinality of transfer event-point set is equal to epPN , that is set T 1, 2, min , epPN . 7.3. Decomposed Models In addition to non-linking variables and constraints present in (P), decomposed models (LBSP) and (LPSP) includes additional constraints and variables to define complicating variables in terms of new transfer event-point t . We first present finished product blending and delivery scheduling problem (LBSP) and then relaxed production unit scheduling model (LPSP). 217 7.3.1. Relaxed Blend Scheduling Problem Product tanks BN1 Kif s , j , k , n Component tanks Blenders Tsf j , k , n , Tss j ,k ,n kCf sB,t kCTf sB,t , kCTssB,t BN2 Cf sB, j ,t B j ,t CTf , CTs Uif s , j ,n B j ,t Tf i , j , n , Tsi , j , n Figure 7.2 Spatial decomposition of blend sub-network structure (BN) Finished product blending and delivery operations correspond to sub-structure BN shown in Figure 7.2 in detail. In this problem we assume that all blend component tanks are dedicated tank and blend components are either supplied directly to a blend unit or are supplied to component tanks. Blending operations features parallel blend units, multiple non-identical blend units producing different products, multipurpose blend units, and multipurpose product tanks. BN can be decomposed into smaller sub-structures using following rule: sub-structures do not have common blend components, products, units, and tanks. Spatial decomposition using this rule is carried out in Figure 7.2 and we obtain two sub-structures BN1 and BN2. Based on this decomposition, problem (LBSP) can be decompose into two independent problems (LBSP1) and (LBSP2), where model (LBSP1) contain all the variables and equations relating to sub-network BN1 and problem (LBSP2) includes all the variables and equations relating to sub-network BN2. For a given refinery network, additional decomposition of BN can be carried out to obtain more 218 than 2 sub-networks and corresponding decomposed scheduling models if possible and set BN includes all the decomposed sub-network of BN, e.g. BN BN1, BN 2 . These independent decomposed scheduling models can be solved using different number of event-points and parameter epbnmin gives minimum number of event-points needed to solve scheduling problem corresponding to sub-network bn to optimality. Since original complicating variables ( JJf s , j , j ,n , Kif s , j ,k ,n , ins , j , k , n , Tss j , k , n , Tsf j , k , n ) are eliminated from (LBSP), we make following changes: - If a blend component is supplied directly to unit without storage options, then the blend component is treated as a raw material for blend unit and flow of component s to blend unit j at event-point n is given by Uif s , j ,n - A unit is added set J for BN to represent a dummy production unit - For blend component that can be stored into tanks, variable Kif s , j ,k ,n represent flow of component s into tank k from dummy production unit j at event-point n The start and finish time ( Tss j , k , n , Tsf j ,k ,n ) for flow into the component tank at event-point n is define by Eq. (1). Kif s , j , k , n M Tsf j , k , n Tss j , k , n M 1 ins , j , k , n , s Sbc , k K s , j J dmy , n N (1) Here, the binary variable ins , j ,k ,n is defined by equation (7a) in Chapter 4 and is equal to 1 if the blend component is flowing into tank at event-point n . If component is supplied to the tank from a dummy production unit j at event-point n , then total amount is bounded by product of big-M and duration of flow into the tank. To represent complicating variables in terms of transfer event-point t instead of eventpoint n , we introduce 0-1 continuous variables x j ,n,t and xks, n,t . These variables are 1 if a transfer event-point t to corresponds to an active event-point n , otherwise 0. 219 x Bj , n ,t wvi , j , n , j J BU J ds , n N , t T , n t iI j x Bj , n ,t 1 x Bj , n ,t 1 x Bj , n ,t , j J BU J ds , n N , t T , n t (2b) x Bj , n,t , j J BU J ds , n N , t T , n t (2c) t t ,t n n n , n t x Bj , n ,t wvi , j , n iI j x n t B j , n ,t (2a) t t ,t n x Bj , n ,t , x Bj , n ,t n n , n t x Bj , n ,t , j J BU J ds , n N , t T , n t j J BU J ds , t T , t T , t t n t (2d) (2e) x Bj , n,t x Bj , n,t 1, j J BU J ds , n N , n N , n n, t T , t T , t t , n t , n t , n t , n t (2f) x Bj , n,t 0, j J BU J ds , n N , t T , n t (2g) Constraints (2a)-(2g) are for a blend unit that receives at least one blend component directly from production unit operations. Binary assignment variable wvi , j ,n is 1 if task i is active at event-point n for unit j . For a blend unit j that is performing some task at event-point n , Eqs. (2a)-(2d) only one event-point t to the event-point n . Eqs. (2e) enforce that event-point t t should be assign an active event-point n before event-point t . More ever, for any two active event-points n n , n must be assigned to an event-point t before assigning n to an event-point t , where t t . Constraints (2a)-(2e) eliminates an occurrence of a situation where x j , n,t 1 and x j , n,t 0 for t t . Since there is one to one correspondence between an active event-point n and event-point t for a unit j , we introduce Eq. (2g) to fix variable x j , n,t 0 whenever n t . Similar to constraints (2a)-(2g) for blend unit that receive components directly from production unit operations, we present Eqs. (3a)-(3g) for blend component tanks as follows: xksB, n,t k K s , jJ dmy xksB, n ,t 1 t t ,t n ins , j , k , n , s Sbc Sk , n N , t T , n t (3a) xk sB, n ,t , s Sbc S k , n N , t T , n t (3b) 220 xk sB, n ,t 1 xksB, n,t xk n t n n , n t jJ dmy B s , n ,t xksB, n,t , s Sbc Sk , n N , t T , n t ins , j , k , n xk sB, n ,t , nt t t ,t n xksB, n,t n n , nt xk sB, n,t , s Sbc Sk , k K s , n N , t T , n t s Sbc S k , t T , t T , t t (3c) (3d) (3e) xksB, n,t xksB, n,t 1, s Sbc Sk , n N , n N , n n, t T , t T , t t , n t , n t , n t , n t (3f) xksB,n,t 0, s Sbc Sk , n N , t T , n t (3g) Complicating variables ( Cf s , j ,t , KCf s ,t , CTss , j ,t , CTf s , j ,t , KCTss ,t , KCTf s ,t ) expressed through transfer event-point t are defined below: Cf sB, j ,t Uif s , j , n M 1 x Bj , n ,t , j J BU , s S cj / S k , n N , t T , n t (4a) Cf sB, j ,t Uif s , j , n M 1 x Bj , n ,t , j J BU , s S cj / S k , n N , t T , n t (4b) Cf sB, j ,t M x Bj , n ,t , (4c) n t j J BU , s S cj / S k , t T CTs Bj ,t Tsi , j , n UH 2 wvi , j , n x Bj , n,t , j J BU J ds , i I j , n N , t T , n t (5a) CTs Bj ,t Tsi , j , n UH 2 wvi , j , n x Bj , n,t , j J BU J ds , i I j , n N , t T , n t (5b) CTf jB,t Tf i , j , n UH 2 wvi , j , n x Bj , n ,t , j J BU J ds , i I j , n N , t T , n t (5c) CTf jB,t Tf i , j , n UH 2 wvi , j , n x Bj , n ,t , j J BU J ds , i I j , n N , t T , n t (5d) CTs Bj ,t UH x Bj , n ,t , j J BU J ds , t T (5e) CTf jB,t UH x Bj , n ,t , j J BU J ds , t T n t nt (5f) Flow of component s from sub-network PN to blend unit j at transfer event-point t is equal to raw material flow into the blend unit at event-point n if x j ,n,t equal to 1 as stated by Eqs. (4a)-(4b). Eq. (50c) forces Cf s , j ,t to be zero if at transfer event point t , no flow exists between PN and a blend unit in BN. As given in Eqs. (5a)-(5d), timing variables for component flows at event-point t into a blend unit j are determined by start and finish time of blend run-mode at 221 corresponding event point n . Eqs. (5e)-(5f) force start and finish time to be zero at event-point t if the event-point t is not assigned to any event-point n . Similarly, he complicating variables associated with blend component tanks are determined by Eqs. (6a)-(7f) as shown below. Aggregated component s stream connecting PN and BN sub-networks is split into multiple streams connecting suitable component tanks as expressed by Eqs. (6a)-(6b). KCf sB,t KCf sB,t KCf sB,t Kif s , j , k , n M 1 xksB, n ,t , s Sbc Sk , n N , t T , n t (6a) Kif s , j , k , n M 1 xksB, n ,t , s Sbc Sk , n N , t T , n t (6b) k K s , jJ dmy k K s , jJ dmy V xk k K s max k n t B s , n ,t , s Sbc S k , t T (6c) KCTssB,t Tss j , k , n UH 2 ins , j , k , n xk sB, n ,t , s Sbc , k K s , j J dmy , n N , t T , n t (7a) KCTssB,t Tss j , k , n UH 2 ins , j , k , n xk sB, n ,t , s Sbc , k K s , j J dmy , n N , t T , n t (7b) KCTf sB,t Tsf j , k , n UH 2 ins , j , k , n xk sB, n ,t , s Sbc , k K s , j J dmy , n N , t T , n t (7c) KCTf sB,t Tsf j , k , n UH 2 ins , j , k , n xk sB, n ,t , s Sbc , k K s , j J dmy , n N , t T , n t (7d) KCTssB,t UH xk sB, n ,t , s Sbc S k , t T (7e) KCTf sB,t UH xk sB, n ,t , s Sbc S k , t T (7f) n t nt If no transfer of flow exists between BN and PN for component s that can be stored in tanks at transfer event-point t , then the corresponding transfer variables are fixed to zero as enforced by Eqs. (6c), (7e) and (7f). 7.3.2. Relaxed Production Unit Scheduling Problem In production unit operations, more than one production unit can produce a blend component s and these blend component can be either directly supplied to a blend unit j J BU or can be sent to a blend component tanks but not both. Original complicating variables ( JJf s , j , j ,n , Kif s , j ,k ,n , ins , j , k , n , Tss j , k , n , Tsf j , k , n ) are eliminated in model (LPSP) and blend components are defined as 222 products instead of intermediates in PN (Figure 7.3). Depending on if the blend component can be stored in tank or not in BN, following changes are made: For blend component s supplied directly to a blend unit j without storage options and - produced by production unit j at event-point n , the flow is represented by Uofbs , j , j , n . - For blend component s that can be stored into tanks and produced by production unit j at event-point n , the flow is given by Uof s , j ,n . Production units PN Uof s , j ,n Tf i , j , n , Tsi , j , n kCf sP,t kCTf sP,t , kCTssP,t Cf sP, j ,t Uofbs , j , j ,n CTf jP,t , CTs Pj ,t Tf i , j , n , Tsi , j , n Figure 7.3 Production unit operations network The material balance constraints (3a) in section 4.3 for blend component s and for production unit j and valid inequality (42) are updated as to define blend components as raw material for PN as follows: bp iI j s ,i , j , n k K kp j Ks k K jpk Kif s , j , k , n JJf s , j , j , n Uof s , j , n j J c j J seq j Js Ks Kof s , k , j , n BU Uofbs , j , j ,n , s S , j J sp , n N J sc JJf s , j , j , n Uif s , j , n p j J seq j Js (3a‟) k K jpk K s Kif s , j , k , n c j J seq j Js JJf s , j , j , n Uof s , j , n Uofbs , j , j , n , j J , n N (42‟) j J BU J sc To determine if transfer happening between production unit j and blend unit j at event-point n , a binary variable a j , j,n is introduced. This variable is needed because refinery might have 223 multiple blend units that can consume blend component s . According to Eq. (8), if production unit j is supplying a component to a blend unit j at event-point n then binary variable a j , j,n is equal to 1. max Uofbs , j , j , n bsmax s Sc / Sk , j J sp , j J bu J sc , n N , j RB j UH a j , j , n , (8) Here, bsmax is maximum blend recipe component s and RBmax is maximum blending ,j j rate for blend unit j and they are determined using the available information about blending operations. Eqs. (9a)-(9d) determines maximum/minimum blend recipe for components and maximum/minimum production rate for each blend units, respectively. c ,max bsmax , s Sbc , j J BU J sc , j max s , i (9a) c ,min bsmin , s Sbc , j J BU J sc , j min s , i (9b) RB max max Rimax j J BU j , j , (9c) RB min min Rimin j J BU j , j , (9d) iI j iI j iI j iI j The flow of a component into sub-network BN is bounded by total storage capacity in BN at any given time as stated by Eq. (10). Uof s , j , n V k K s max k wvi , j , n , s Sbc Sk , j J sp , n N iI j I sp (10) To obtain the complicating variables in terms of transfer event-point t , constraints similar to (2a)-(2g) are added to production scheduling problem. For a blend unit that receives at least one blend component directly from PN, Eqs. (11a)-(11g) are used and for a blend component that has storage options available in BN, Eqs. (12a)-(12g) are used to determine active transfer flow at event-point t . Here, variables x j ,n,t and xks,n,t are continuous 0-1 variables. x Pj , n,t j J seq j a j , j , n , j J BU J ds , n N , t T , n t (11a) 224 x Pj , n ,t 1 x Pj , n ,t 1 x Pj , n ,t , j J BU J ds , n N , t T , n t (11b) x Pj , n,t , j J BU J ds , n N , t T , n t (11c) (11d) t t ,t n n n , n t x Pj , n ,t a j , j , n x n t P j , n ,t t t ,t n x Pj , n ,t , x Pj , n ,t n n , nt x Pj , n,t , j J BU J ds , j J seq j , n N,t T , n t j J BU J ds , t T , t T , t t n t (11e) x Pj , n,t x Pj , n,t 1, j J BU J ds , n N , n N , n n, t T , t T , t t , n t , n t , n t , n t (11f) x Pj , n,t 0, j J BU J ds , n N , t T , n t (11g) xksP, n,t wvi , j , n , s Sbc Sk , n N , t T , n t iI sp , jJ i xksP, n ,t 1 xk sP, n ,t 1 xksP, n,t xk sP, n ,t , s Sbc S k , n N , t T , n t (12b) xksP, n,t , s Sbc Sk , n N , t T , n t (12c) t t ,t n n n , n t wv i, j ,n iI sp xk n t P s , n ,t (12a) t t ,t n xk sP, n ,t , nt xksP, n,t n n , nt xksP, n,t , s Sbc S k , j J sp , n N , t T , n t s Sbc S k , t T , t T , t t (12d) (12e) xksP, n,t xksP, n,t 1, s Sbc Sk , n N , n N , n n, t T , t T , t t , n t , n t , n t , n t (12f) xksP,n,t 0, s Sbc Sk , n N , t T , n t (12g) Similar to Eqs. (4a)-(7f) used to determine flow amounts and duration of components flow into BN, Eqs. (13a)-(16f) are used to determine the complicating variables associated with transfer event as follows: Cf sP, j ,t Uofb max P bsmax j J BU , s S cj / S k , n N , t T , n t , j RB j UH 1 x j , n , t , (13a) Uofb max P bsmax j J BU , s S cj / S k , n N , t T , n t , j RB j UH 1 x j , n , t , (13b) s , j , j , n j J sp Cf sP, j ,t s , j , j , n j J sp max Cf sP, j ,t bsmax UH x Pj , n ,t , , j RB j nt j J BU , s S cj / S k , t T (13c) 225 CTs Pj ,t Tsi , j ,n UH 3 wvi , j ,n a j , j ,n x jP,n ,t , j J BU J ds , j J jseq ,i I j ,n N ,t T ,n t (14a) CTs Pj ,t Tsi , j ,n UH 3 wvi , j ,n a j , j ,n x jP,n ,t , j J BU J ds , j J jseq ,i I j ,n N ,t T ,n t (14b) CTf jP,t Tf i , j ,n UH 3 wvi , j ,n a j , j ,n x jP,n ,t , j J BU J ds , j J jseq ,i I j ,n N ,t T ,n t (14c) CTf jP,t Tf i , j ,n UH 3 wvi , j ,n a j , j ,n x jP,n ,t , j J BU J ds , j J jseq ,i I j ,n N ,t T ,n t (14d) CTf jP,t UH x Pj , n, t , j J BU J ds , t T (14e) CTf jP,t UH x Pj , n, t , j J BU J ds , t T nt nt KCf sP,t Uof s, j, n KCf sP,t Uof s, j, n jJ sp KCf sP,t V 1 xksp, n, t , s Sbc Sk , n N , t T , n t (15a) V 1 xksp, n, t , s Sbc Sk , n N , t T , n t (15b) max k n t max k k K s V xk k K s max k k K s jJ sp (14f) P s , n, t , s Sbc S k , t T (15c) KCTs sP,t Tsi , j ,n UH 2 wvi , j ,n xk sP,n ,t , s S bc , S k ,i I sp , j J i ,n N ,t T ,n t (16a) KCTs sP,t Tsi , j ,n UH 2 wvi , j ,n xk sP,n ,t , s S bc , S k ,i I sp , j J i ,n N ,t T ,n t (16b) KCTf sP,t Tf i , j ,n UH 2 wvi , j ,n xk sP,n ,t , s S bc , S k ,i I sp , j J i ,n N ,t T ,n t (16c) KCTf sP,t Tf i , j ,n UH 2 wvi , j ,n xk sP,n ,t , s S bc , S k ,i I sp , j J i ,n N ,t T ,n t (16d) KCTssP,t UH xk sP, n, t , s Sbc S k , t T (16e) KCTf sP,t UH xk sP, n, t , s Sbc S k , t T (16f) n t nt As mention before, the multiple streams for a blend component are aggregated into one stream to reduce complicating variables in augmented Lagrangian decomposition algorithm. Thus, the flow of component s between PN and BN at transfer event-point t is determined in Eqs. (13a), (13b), (15a), and (15b) by adding component s flow from multiple production units. More ever, flow max of component s to a blend-unit j is bounded by maximum consumption rate ( bsmax , j RB j UH ) 226 and flow to storage tanks is limited by maximum available storage capacity ( Vsmax ). Duration k K s of a transfer event between PN and a blend unit j is determined by the start and finish time of run-modes of production units that are supplying material to the blend unit (i.e. a j , j ,n = 1). In addition to determine the timing variables for a direct flow of components between a production unit and a blend unit at transfer event-point t , Eqs. (14a)-(14d) enforces that same start and finish time for flow of all components that can be consumed by the blend unit. Similarly, the (16a)(16d) assigns start and finish time for flow of components that can be stored in tanks. If no flow transfer between PN and BN at event-point t , then corresponding linking variables are zero as enforced by Eqs. (13c),(14e),(14f), (15c), (16e), and (16f). 7.4. Strengthening the Augmented Lagrangian Decomposition When refinery operations network is decomposed, information regarding the operational and logistics rules that govern flow of components between PN and BN is lost. To incorporate many of these rules without destroying separability, additional constraints are included in (LBSP) and (LPSP) and presented in section 7.4.2. These constraints speed up convergence to a feasible solution in augmented Lagrangian optimization algorithm and in most instances push feasible solution closer to a global optimal solution. To further improve performance of augmented Lagrangian optimization, valid inequalities can be added to improve computational efficiency of decomposed models and they are presented in Appendix Chapter 7 for model (LBSP) and (LPSP). To improve integration between PN and BN, production capacity limitation of production unit operations should be obtained before start of the optimization algorithm. 7.4.1. Preprocessing Step When refiner network is decomposed into (BN) and (PN), the resulting decomposed scheduling problem (LBSP) does not take into consideration the production capacity limitation of 227 production unit operations. Therefore, a pre-processing step before the start of the algorithm is needed to determine maximum and minimum production rate of each blend component. Relaxed production unit operation scheduling model (LPSP) is optimized for one of the following objective function to obtain maximum and minimum production rate for a blend component s . z Cf s , j ,T CTf j ,T CTs j ,T , s Sbc \ S k , j J sc (17a) z KCf s ,T KCTf s ,T KCTss ,T , s Sbc S k (17b) Objective function (17a) is maximized to determine the maximum production flow rate of a component s to a blend unit j and Eq. (17b) is used to obtain the maximum production rate for components that have storage options available. The minimum flow rate would always be zero unless constraint (18a)-(18b) is added to the model to require flow to happen. x j , n ,T 1, s Sbc \ S k , j J sc (18a) n xk s , n ,T 1, s Sbc S k (18b) n Eq. (18a) requires the flow of component s from PN to a blend unit at transfer event-point T and for component that can be stored in tanks, Eq. (18b) enforces flow to happen at event-point T . For blend units that receive all the components directly from production unit operations without any storage, the maximum and minimum total flow rate into a blend unit is determined using objective function given in Eq. (19). z Cf sSbc s , j ,T CTf j ,T CTs j ,T , j J BU J ads (19) P Upon solving the pre-processing problems, we obtain values of variables { Cf s , j ,t , CTs Pj,t , CTf jP,t , x Pj, n ,t , KCf sP,t , KCTssP,t , KCTf sP,t , xk sP, n ,t } from the optimal solution. RPus , j Cf sP, j ,T x Pj , n,T n Tf jP,T Ts Pj ,T , j J BU , s S cj \ Sk (20a) 228 RPks KCf sP, j ,T xksP, n,T n KTf sP,T KTssP,T Cf x P s , j ,T RPj sS cj , s Sbc Sk (20b) , j J BU J ads (20c) P j , n ,T n Tf jP,T Ts Pj ,T Flow rate for a component from PN to a blend unit is calculated using Eq. (20a), total flow rate into the blend unit is determined using Eq. (20c) and Eq. (20b) is used for components that are stored in tanks. 7.4.2. Strengthening Decomposed Problems Many constraints that are unnecessary or redundant in full-scale model (P) are needed in decomposed model to obtain a solution that is much more align with a feasible solution of fullscale problem and to speed up convergence to a feasible solution in augmented Lagrangian algorithm. In this section, we propose constraints that improve performance of the decomposition algorithm. 7.4.2.1. Relaxed Blend Scheduling Problem Lower bound on the duration of the component flow from PN to BN is determined by minimum run-length ( RL ) requirement for production units in PN. Thus, constraints (21b) similar to Eq. (14) of original model in section 4.3 is included in relaxed problem. Tsf j , k , n Tss j , k , n RL ins , j , k , n UH 1 ins , j , k , n , s Sbc , k K s , j J dmy , n N (21b) For a case when more than one blend component can be stored in component tanks in BN and if these blend components are produced by production units in PN that are interconnected, that is these production units must operate at the same time, then components produced by these units would be supplied to BN at the same time. ins , j , k , n in k K s s , j , k , n , s Sbc , s S bc , k K s , j J dmy , n N , cont s , s 0 (22) 229 k K s , k K s ins , j , k , n in k K s s , j , k , n , s Sbc , s S bc , j J dmy , n N , conts , s 0 (23) Here, parameter conts , s is equal to 1 if components s and s are produced by units in PN that are interconnected. Eqs. (22)-(23) states that if a blend unit s is being supplied to a component tank at event-point n then component s must also be supplied to at least one suitable component tank at event-point n. Timing constraints for these connected blend components are given in Eqs. (24a)(24d) which enforce same start and finish time for component flow at event-point n. Tss j , k , n Tss j , k , n UH 2 ins , j , k , n ins , j , k , n , s Sbc , s Sbc , k K s , k K s , j J dmy , n N , conts , s 0 Tss j , k , n Tss j , k , n UH 2 ins , j , k , n ins , j , k , n , s Sbc , s Sbc , k K s , k K s , j J dmy , n N , conts , s 0 Tsf j , k , n Tsf j , k , n UH 2 ins , j , k , n ins, j , k , n , s Sbc , s Sbc , k K s , k K s , j J dmy , n N , conts , s 0 Tsf j , k , n Tsf j , k , n UH 2 ins , j , k , n ins, j , k , n , s Sbc , s Sbc , k K s , k K s , j J dmy , n N , conts , s 0 (24a) (24b) (24c) (24d) Furthermore, if multiple tanks are available to store a blend component, then following sequence constraint (25) are needed as it captures what happens in during production unit operations in PN. Tss j ,k ,n 1 Tsf j ,k ,n , s Sbc , k Ks , k Ks , k k , j J dmy , n N (25) Set J ds includes all blend units that receive at least one blend component directly from production unit operations without any storage. If all the components consumed by blend unit j are supplied directly without storage, then that blend unit belongs to set J ads . The production capacity of blend components determined in pre-processing step is incorporated into relaxed blend scheduling by removing the big-M term from model. Kif s , j , k , n RPk smax Tsf j , k , n Tss j , k , n RPk smaxUH 1 ins , j , k , n , s Sbc , k K s , j J dmy , n N B Cf sB, j ,t Uif s , j , n RPusmax j J BU , s S cj / S k , n N , t T , n t , j UH 1 x j , n , t , (1‟) (4a‟) 230 B Cf sB, j ,t Uif s , j , n RPusmax j J BU , s S cj / S k , n N , t T , n t , j UH 1 x j , n , t , (4b‟) B Cf sB, j ,t RPusmax , j UH x j , n , t , (4c‟) n t KCf sB,t KCf sB,t j J BU , s S cj / S k , t T Kif s , j , k , n RPksmaxUH 1 xk sB, n ,t , s Sbc Sk , n N , t T , n t (6a‟) Kif s , j , k , n RPksmaxUH 1 xksB, n ,t , s Sbc Sk , n N , t T , n t (6b‟) k K s , jJ dmy k K s , jJ dmy Eqs. (1), (4a), (4b), (4c), (6a), and (6b) are replaced with updated constraints (1‟), (4a‟), (4b‟), (4c‟), (6a‟), and (6b‟). Upper and lower limit on flow into BN from PN at transfer event-point t is given by Eqs. (26a)-(28b). B B max B Cf sB, j ,t RPusmax j J BU , s S cj / S k , t T , j CTf j , t CTs j , t RPu s , j UH 1 x j , n ,t , n t (26a) B B max B Cf sB, j ,t RPusmin j J BU , s S cj / S k , t T , j CTf j , t CTs j , t RPu s , j UH 1 x j , n ,t , n t (26b) KCf sB,t RPksmax KCTf sB,t KCTssB,t RPk smaxUH 1 xk sB, n,t , s Sbc S k , t T n t (27a) KCf sB,t RPksmin KCTf sB,t KCTssB,t RPk smaxUH 1 xk sB, n ,t , s Sbc S k , t T nt (27b) Cf sB, j ,t RPjmax CTf jB,t CTs Bj ,t RPjmaxUH 1 x Bj , n,t , j J BU J ads , t T n t / Sk (28a) Cf sB, j ,t RPjmin CTf jB,t CTs Bj ,t RPjmaxUH 1 x Bj , n ,t , j J BU J ads , t T n t / Sk (28b) sS cj sS cj Flow of component s to blend unit j is bounded by product of duration of flow and min maximum/minimum production rate ( RPu max s , j / RPu s , j ) as limited by Eqs. (26a) and (26b), respectively. If no flow exists between PN and a blend unit (i.e. x j , n ,t 0 ) then inequality in n (26a)-(26b) is relaxed. For a blend component that can be stored in tanks, if no flow exists between PN and BN for a (i.e. xk n s , n ,t 0 ) then inequalities (27a)-(27b) are relaxed. Similarly, 231 for blend units that receive all their components directly from production unit operations, the total flow amount is bounded by upper and lower limits as enforced by Eqs. (28a)-(28b). A lower bound of (LBSP) is obtained, when the model (LBSP) is solved to optimality with all the Lagrangian and augmented penalty parameter fixed to zero. Thus, solution f BSP ( m 0) at algorithm count m 0 gives the best lower bound of (LBSP) and Eq. (29) is added to model. f BSP f BSP ( m 0) , if m 1 (29) For a refinery that has multiple dedicated tanks for blend component storage, Eq. (30) is added to model when algorithm iteration count is m 1 . in n N s, j ,k ,n 1, s Sbc , k K sh , j J dmy , m 1 if in ( m 0) s, j ,k ,n 0 n (30) The blend scheduling problem become difficult to solve as the additional flexibility of a parallel blend component tanks is introduced. To reduce this complexity, Eq. (30) is added for parallel blend component tanks. In our work, since we do not maintain minimum inventory levels for components or blend products, following constraints can be added to simplify the blend scheduling problem. If one were to require either minimum inventory level for blend components but not finished blend product, then constraint (31) can be added to the model without compromising the optimal solution. Finished product blend tanks cannot have simultaneous loading and unloading operations for security reasons and to allow time for mixing and product specification certificate analysis. For this reason, if a blend unit is receiving all of its blend components from storage tanks, then this blend unit would not be active during last event point N because the product produce will not contribute to demand orders fulfillment and there is no minimum inventory requirement for products. We add Eqs. (31) to (LBSP) to fix blend unit allocation wvi , j , N to zero. wvi , j ,n 0, j J ads , i I j , n N (31) 232 Furthermore, linking variables are compared for transfer event-point t and not event-point n , the scheduling problem corresponding to sub-network PN can be solved using less number of eventmin min min points n ( epPN ) than scheduling problem corresponding to sub-network BN. When epPN , epbn then following variables are zero at last event-point in (LBSP) when no minimum inventory level requirement for components and blend products. min ins, j , k , n 0, s Sbc , k Ks , j J dmy , n N , if epPN epbnmin (32a) min min ins, j , k , n 0, j J BU J ds , s S jp , k Ks , n N , if epPN epbn (32b) min min wvi , j ,n 0, j J BU J ds , i I j , n N , if epPN epbn (32c) min min x j , n,t 0, j J BU J ds , t T , n N , if epPN epbn (32d) min xks, n,t 0, s Sbc , t T , n N , if epPN epbnmin (32e) min Cardinality of set T epPN and transfer event-point t only corresponds to an event-point n t . min min When epPN , flow of components into the BN happens only to satisfy finished product epbn demand, and Eqs. (32a)-(32e) fixes some binary and 0-1 continuous variables to zero for n N . If one were to require minimum inventory level for products and components, then these constraints should not be included in the model. 7.4.2.2. Relaxed Production unit Scheduling Problem To insure that demand commitments for finished blend product are met and excess inventory is minimized in BN, Eqs. (33a) –(33b) are included when parameter is r not zero. The parameter r is different than algorithm iteration count m and will be defined in section 7.5. min sSbc Sk , jJ sp , n epbn Uof s , j , n Uofbs , j , j , n min sSbc / Sk , jJ sp , j J BU J sc , n epbn sS A S bn , oOs Do, s sS bn , k K s stos , k DmPbn , bn BN , r 1 (33a) 233 Uof s , j , n Uofbs , j , j , n min sSbc / S k , jJ sp , j J BU J sc , n epbn min sSbc Sk , jJ sp , n epbn (33b) Do, s InvtPbn , bn BN , r 1 sS A S bn , oOs Each independent sub-networks of BN has their own set of finished blend product demand orders and initial storage inventory. Eq. (33a) requires production unit operation to produce at least minimum amount of total components to satisfy demand for each blending operations subnetwork. More ever, if the total demand cannot be met, then a positive slack variable DmPbn is added to facilitate feasible solution. Eq. (33b) limits the total production of blend components as it is bounded by maximum demand for finished blend products. For cases where production exceeds, a positive slack variable InvtPbn is introduced and is later penalized in objective function. To ensure that demand is met on time and to improve performance of model, constraints (34a) and (35c) are added to the model when parameter is r 1 . CTs Pj ,t CTs j ,t esPj ,t , bn BN , j J bnBU , t T , t epbnmin , Cf s , j ,t 0, r 1 (34a) CTs Pj ,t CTs j ,t esPj ,t , bn BN , j J bnBU , t T , t epbnmin , Cf s , j ,t 0, r 1 (34b) CTf jP,t max lePj , bn BN , j J bnBU , t T , t epbnmin , Cf s , j ,t 0, r 1 min (34c) B B s B B s B t epbn s B B B B min KCTssP,t KCTs s ,t eskPs ,t , bn BN , s Sbc , t T , t epbn , KCf s ,t 0, r 1 (35a) min KCTssP,t KCTs s ,t eskPs ,t , bn BN , s Sbc , t T , t epbn , KCf s ,t 0, r 1 B (35b) B min KCTf sP,t max KCTf s ,t lekPs , bn BN , s Sbc , t T , t epbn , KCf s ,t 0, r 1 min t epbn (35c) Constraints (34a)-(34b) and (35a)-(35b) encourage the start and finish time of flow to be near the corresponding timing variables‟ solution obtain solving (LBSP). Here esPj ,t , eskPs,t , lePj , and lekPs are positive slack variables which are penalized in objective function. Constraint (36) requires the objective function to be bounded by f PSP (LBSP). r 0 , similar to Eq. (73) in 234 f PSP f PSP ( r 0) r 1 , if (36) x Pj , n ,t 1, j J BU , t T , Cf s , j ,t 0, r 1 xksP, n ,t 1, s Sbc Sk , t T , KCf s ,t 0, r 1 nN , n t nN , n t B (37a) s B (37b) Since PN is producing blend components for BN, we add following requirement (37a)-(37b) in model (LBSP) when flow variables are nonzero in BN for transfer event-point t . The slack variables ( DmPbn , Invtbn , esPj ,t , lePj , eskPs ,t , lekPs ) are penalized and the updated objective function for the model (LPSP) is given below: min ff PSP z s , j ,t sSbc , jJ BU sSbc , t Cf sP, j ,t ,t s ,t jJ BU j ,t CTs Pj ,t j ,t CTf jP,t ,t KCf s ,t KCTs s ,t KCTf sP,t P s ,t P s ,t Cf Cf CTs CTs CTs CTs CTf CTf CTf CTf Kif KCf KCf Kif KCTs KCTs KCTs KCTs KCTf KCTf KCTf KCTf B P Cf s , j ,t Cf s , j ,t sSbc , jJ BU , t 2 B j ,t 2 P j ,t j ,t 2 B P s ,t s ,t B P s , j ,t s , j ,t B P j ,t j ,t P s ,t B P s ,t sSbc , t B P s , j , k , n 2 P 2 s ,t B s ,t 2 s ,t 2 s , j , k , n 2 B s ,t 2 P 2 j ,t B j ,t P j ,t B jJ BU , t 2 B P s ,t s ,t 2 Uof s , j , n Uofbs , j , j , n Invtbn min min bnBN , r 1 sSbc S k , jJ sp , n epbn sSbc / S k , jJ sp , j J BU J sc , n epbn 11 16 17 18 c DmPbn c esPj ,t c eskPs ,t c lePj c3 bnBN , r 1 c19 sSbc S k ,t , r 1 jJ BU , t , r 1 sSbc Sk , r 1 lekPs j J h min , t epbn , Cf B s , j ,t 0, jJ h , s min jJ h , t epbn , Cf 0, r 1 j J c 20 x j , n ,t B s , j ,t 0, r 1 s c1j x j , n ,t B s , j ,t Cf jJ BU , r 1 h c1j x j , n ,t min , t epbn , r 1 jJ h , s Objective function ff PSP includes original objective function, augmented Lagrangian terms and additional terms that tries to take into consideration the cost that will be incurred in blending 235 operations by minimizing excess inventory for blend components, under production of blend components and component flow start and finish time violations based on solution of LBSP. This transfer time violation is not penalized heavily as it would be the goal of augmented Lagrangian optimization algorithm to eliminate this violation in order to obtain feasible solution of global problem. The last three terms in the objective function is for parallel blend units. If the material flow is happening to at least one parallel blend unit in BN ( Cf s , j ,t 0 ) at event-point t then the B s rest of the parallel units that have no flow happening ( Cf s , j ,t 0 ) at that event-point are B s penalized less heavily than if no flow was happening at any parallel blend unit ( Cf s , j ,t 0 ) at B s, j event-point t . Cost coefficients c16 , c17 , c18 , c19 , and c20 need to be determined along with Lagrangian multipliers and augmented penalty parameters. The coefficients c20 should be 0.50 c20 c1 and parameters c16 , c17 , c18 , c19 should be within [0,1]. 7.5. Augmented Lagrangian Optimization Algorithm Before proposed augmented Lagrangian optimization algorithm is presented, models under for the algorithm are as follows: min f BSP z sSbc , jJ s , j ,t Cf sB, j ,t BU sSbc , t j J ,t s ,t BU CTs Bj ,t j ,t CTf jB,t ,t B s ,t P Cf Cf sB, j ,t Cf s , j ,t P B KCf KCf s ,t KCf s ,t sSbc , t j ,t KCf s ,t KCTssB,t s ,t KCTf sB,t sSbc , jJ BU , t 2 CTf CTf KCTf KCTf P B CTs CTs j ,t CTs j ,t BU jJ , t KCTs 2 KCTs B s ,t P KCTs s ,t 2 CTf B j ,t 2 KCTf B s ,t P 2 j ,t P 2 s ,t s.t. constraints corresponding to units and connections in BN (LBSP) and valid inequalities in Appendix Chapter 7 236 min ff PSP z s , j ,t sSbc , jJ BU , t sSbc , t s ,t Cf sP, j ,t j ,t jJ BU ,t CTs Pj ,t j ,t CTf jP,t KCf s ,t KCTssP,t s ,t KCTf sP,t P s ,t Cf Cf CTs CTs CTs CTs CTf CTf CTf CTf Kif KCf KCf Kif KCTs KCTs KCTs KCTs KCTf KCTf KCTf KCTf B Cf Cf s , j ,t Cf sP, j ,t 2 2 B P s , j ,t s , j ,t sSbc , jJ BU ,t 2 B j ,t CTs 2 B jJ BU , t CTf P j ,t j ,t 2 B KCf sSbc , t P s ,t s ,t KCTs KCTf c3 B s ,t P s ,t B P s ,t s ,t P 2 j ,t B j ,t P j ,t 2 B P j ,t j ,t 2 B P s , j , k , n s , j , k , n 2 B s ,t P 2 s ,t 2 B P s ,t s ,t 2 Uof s , j , n Uofbs , j , j , n Invtbn min min bnBN , r 1 sSbc Sk , jJ sp , n epbn sSbc / Sk , jJ sp , j J BU J sc , n epbn c11 bnBN , r 1 jJ min , t epbn , esPj ,t c17 jJ BU , t , r 1 h DmPbn c16 Cf B s , j ,t 0, jJ h , s Cf s sSbc Sk , t , r 1 eskPs ,t c18 c 20 x j , n ,t B s , j ,t 0, r 1 j J h min , t epbn , lePj c19 jJ BU , r 1 c1j x j , n ,t Cf B s , j ,t 0, r 1 j J h sSbc S k , r 1 lekPs c1j x j , n,t min , t epbn , r 1 jJ h , s s.t. constraints corresponding to units and connections in PN (LPSP) and valid inequalities in Appendix Chapter 7 The Model (LBSP) includes equations, objective function and variables present in model (LRP) corresponding to units and connection in sub-network BN and constraints presented in section 7.3.1, section 7.4.2.1. If sub-network BN can be decomposed into smaller structures, then the model (LBSP) will be decomposed such that model (LBSP1) would only include constraints regarding units and connections present in sub-structure BN1, and so on. Meanwhile, the model (LPSP) includes constraints and variables present in (LRP) pertaining to sub-network PN, constraints proposed in sections 7.3.2 and 7.4.2.2, and objective function ff PSP . Due to quadratic terms in the objective function, models (LBSP) and (LPSP) are mixed integer quadratic programming (MIQP) problems. 237 max zs , j Cf s , j ,T CTf j ,T CTs j ,T , s Sbc \ S k , j J sc s.t. constraints corresponding to units and connections in PN (PCs-1) min zs , j Cf s , j ,T CTf j ,T CTs j ,T , s Sbc \ S k , j J sc s.t. constraints and variables corresponding model (LPSP) and (PCs-2) equation (18a) Models (PCs-1) and (PCs-2) is used to determine maximum and minimum production rate of blend component s to blend unit j , respectively. max zs KCf s ,T KCTf s ,T KCTss ,T , s S bc S k s.t. constraints and variables corresponding model (LPSP) (PKs-1) min zs KCf s ,T KCTf s ,T KCTss ,T , s Sbc S k s.t. constraints and variables corresponding model (LPSP) and (PKs-2) equation (18b) The upper and lower limit of production rate for blend components that are stored is obtained solving models (PKs-1) and (PKs-2), respectively, and these problems are optimized for each blend component. max z j Cf sSbc s , j ,T CTf j ,T CTs j ,T , j J BU J ads s.t. constraints and variables corresponding model (LPSP) (PJ-1) 238 min z j Cf sSbc s , j ,T CTf j ,T CTs j ,T , j J BU J ads s.t. constraints and variables corresponding model (LPSP) and equation (18a) (PJ-2) Maximum and minimum production rate limits for a blend unit that receives all of its components straight from production units is determined using models (PJ-1) and (PJ-2), respectively. All the pre-processing MILP models are independent and can be solved in parallel. Proposed Augmented Lagrangian optimization algorithm is shown in Figure 7.4 and the detailed steps and termination criteria are as follows: Step 1: Solve preprocessing models (PJ-1), (PJ-2), (PCs-1), (PCs-2), (PKs-1), and (PKs-2) in min max parallel. Calculate RPusmax , RPksmin , RPjmax , and RPjmin . , j , RPus , j , RPks Step 2: Set m 0 , r 0 . Initialize Lagrangian multipliers s(0) , s(0),j ,t 0 , s(0),j ,t 0 , s(0),t 0 , , j ,t 0 (0) (0) (0) (0) (0) (0) s(0) s(0) ,t 0 and penalty parameters Cf , CTs , CTf , KCf , KCTs , KCTf . Pick algorithm ,t 0 , parameters , , 1 , and maximum algorithm count M . Step 3: For r 0 , solve model (LBSP) and (LPSP) in parallel and obtain lower bounds f BSP ( m 0) , f PSP ( r 0) , and update r r 1 . Otherwise if r 1 , solve (LBSP). B B Obtain Cf s , j ,t , CTs j ,t , CTf B j ,t P B B B , KCf s ,t , KCTs s ,t , KCTf s ,t . P Step 4: Solve (LPSP) and obtain Cf s , j ,t , CTs j ,t , CTf P j ,t P P P , KCf s ,t , KCTs s ,t , KCTf s ,t . Step 5: Check if violation matrix norm ( g ) meets tolerance . Where, 239 Cf sB, j ,t Cf sP, j ,t B P KCf s ,t KCf s ,t B P CTs j ,t CTs j ,t g B P CTf j ,t CTf j ,t KCTs B KCTs P s ,t s ,t KCTf B KCTf P s ,t s ,t If g , set m m 1 . Update multipliers using equation ( m 1) ( m ) m x xx . m If g m g m 1 , update m1 m and go to step 3. Step 7: If g , a feasible solution is found and the objective function for refinery operation is calculated as follows: f f PSP f bnBN BSP bn H Minimization of time horizon is included in (LPSP) and (LBSP) objective function thus makespan H is subtracted from solution of (LBSP). Step 8: if m M , terminate the algorithm as the maximum iterations limit is reached. Otherwise, go to step 3. 240 Maximum/minimum production Rate for blend components and blend units Solve Pre Processing models Initialize and Lagrange multipliers, penalty parameters Solve in Parallel Solve (LBSP1) Solve (LBSP2) Solve (LBSPn) If Solve (LBSP1) Solve (LPSP) Loop converged? No . Update multipliers and penalty parameters Yes Feasible solution Figure 7.4. Augmented Lagrangian optimization algorithm 7.6. Examples Three different refinery configurations are used to demonstrate the effectiveness of the proposed algorithm in solving scheduling problems. These refinery configurations are: (1) no 241 blend component tanks, (2) storage options for all blend components, and (3) component tanks for some blend unit. A real case study provided by Honeywell Process Solution (HPS) presented in Chapter 4 is used to obtain four different examples. The original Honeywell refinery network includes two non-identical blend unit, parallel crude distillation units (CDUs), 13 other production units, 5 intermediate tanks, 9 product tanks, and 2 charging tanks for one of the CDU while other CDU receives crude oil mixture from pipeline. The refinery produces middle distillate products using 8 production units and four finished blend products using two non-identical blend units. A multipurpose Diesel blender produces three different grades of diesel while jet fuel is blended by jet blender. For each blend products, 2 dedicated product tanks are available. In addition to these tanks, a multipurpose product tank is present that services two different grades of diesel product. A minimum of 6 hours maintenance downtime is required when swing tank switches from servicing lower grade product to high grade. This downtime is essential to remove any sulfur or other contamination from swing tank before low sulfur grade product is serviced. Fill-draw-delay for finished blend product tanks is set to 4 hours to let product to settle down and mix properly for certificate analysis. For all production and blend units, minimum of 6 hours uptime (minimum run-length) is required and total available scheduling horizon is 240 hours. Data for the case study is given in Appendix Chapter 3. Example 1. No component tanks The case study provided by Honeywell is used. Example 2. No component tanks and parallel blend units The original case study (Figure 4.2) is modified by adding a second diesel blend unit alongside the original diesel blender. This new diesel blender produces same three grade products as the original blender and 7 product tanks are available to service diesel blend products produced by both diesel blenders. Table A7-1 gives production rate capacity information about parallel diesel blender. 242 Example 3. Storage tanks for all blend components The case study presented in Figure 4.2 is modified by severing connections between production units and blend units (diesel and jet blenders) adding dedicated storage tanks for 6 different blend components and storage capacity data are provided in Table A7-2. Example 4. Blend component storage and direct flow of component to blend unit In this example, the refinery has both component storage options and direct flow of component from production units to a blend unit. We include 3 dedicated component tanks for the diesel blender and no blend component storage options for the jet blend unit. Decomposition: We decompose the refinery network presented in Figure 4.2 into sub-networks PN and BN by splitting pipeline that transports components from a production unit to a blend unit or component storage tank. Thus, sub-network PN contains charging tanks, CDUs, production units, intermediate tanks and produces blend components, intermediates and intermediate distillate products that are supplied directly to market without storage options whereas sub-network BN includes diesel and jet blenders, component tanks (if any), product tanks, and produces finished blend products diesel and jet fuel. These two different blend products are not stored in same tanks neither do diesel and jet blenders share a common blend component. Thus, we can further decompose sub-network BN into two sub-networks, BN1 and BN2. Network BN1 containing diesel blender, tanks for components blended in diesel blender (if any), and product tanks that service diesel products and corresponding model is (LBSP1). Whereas, sub-network BN2 includes jet blender, components tanks and jet fuel product tanks and relaxed model is (LBSP2). For example 2, two parallel diesel blenders would be present in one network because two parallel blenders share blend components and produce diesel fuels that are stored in same product tanks. 243 7.6.1. Computational Results In this section, we investigate the performance of the proposed augmented Lagrangian optimization algorithm. The algorithm is evaluated with three performance measures: number of iterations, and computational time (seconds), and quality of the solution. Data for cost coefficients parameters for objective function is given in Table 7.1. Table A7-3 gives demand for intermediate distillate products. Table A7-4 provides demand for finished blend products, products CARB diesel, EPA diesel, Red-dye diesel, and Jet fuel are referred to as product P1, P2, P3 and P4, respectively. Intermediate demand orders are bound by maximum and minimum amount and can be delivered anytime during the delivery window. Maximum delivery rate, product unloading rate to satisfy demand, is 10 (kbbl/h). All the optimization models are formulated using GAMS, decomposition algorithm implemented in MATLAB R2014a and solved on a Dell® Optiplex 9020 computer with 2.90 GHz Quad Core Intel® Processor i5-4570S, CPU 2.90GHz, and 16.0 GB memory. CPLEX 12.3.0/GAMS 23.7.1 solver is used to solve both MILP and MIQP problems. Optimality tolerance of 1e6 % is used for full-scale and preprocessing step models, while for (LBSP) and (LPSP), optimality tolerance of 1% is used and for the algorithm, maximum number of iteration is set to 100. Table 7.1 Parameters in objective function Parameter Value*100 1 CCarbNormal,DieselBlender 10 Parameter Value*100 Cs7, s 40 (Unfavorable: 60) C1EPA Normal,DieselBlender 7 Cs8, s 40 (unfavorable: 60) C1ReddyeNormal,DieselBlender 4 Ck9 5 C1JetNormal,Jet Blender 5 C10 s,k 5 244 Ck2 0.51 (Swing tank: 1.5) 50 Vkmax Ck3 1 Co11 1150 Co12 800 C 4j 50 C13 s 1100 Ck5 10 (Swing tank: 40) Co14 5400 Ci6,i 100 (Unfavorable: 250) Co15 3250 Example 1. This example involves no blend component tanks, 1 multipurpose diesel blender producing three different grades of diesel products that can be stored in 7 different product tanks and 1 jet blender producing jet fuel that can be stored in 2 tanks. One of the diesel product tanks is a swing tank that can store two different type of diesel fuel. For this case study, we select parameters as c16 UH e 2 and c18 UH e 1 . These set of parameters drives solution of 1 1 model (LPSP) for iteration m 0 to that has linking timing variables values much closer to that obtain by (LBSP). This Model sizes for different set of demand examples are shown in Table 7.2. As the number of demand order increase, size of the model increase too and decomposed models have significant less numbers of variables and constraints compared to full-scale model. Furthermore, the complicating variables are expressed in terms of transfer event points, we are able to solve decomposed model with different set of event-points. Full scale model of problem 1 requires minimum of 6 event points to obtain optimal solution while using proposed approach, we are able to solve relaxed diesel blend scheduling model (LBSP1) is solved using 5 event points, relaxed jet blend scheduling model (LBSP2) using 6, and production unit scheduling model (LPSP) using 5 event points. The total numbers of transfer event-points are always equal to the number of event-points n used to solve (LPSP). In Table 7.3, we compare solutions of full space 245 model and proposed decomposition approach for problems of varying demand requirements. In Table 7.3, column „ g g ‟ represents the value of the augmented and penalty term in the 2 augmented Lagrangian function, „ g ‟ represents norm of the consistency constraints. The computational time to obtain optimal solution increases as the complexity of the model increases and for problem 7, optimality gap is still at 12.91% even after 8 hours. The performance of the algorithm in terms of solution and the quality of the solution is shown in Table 7.3. For all the problems, proposed algorithm provides feasible solution that is closer to the optimal solution reported by full-scale model. For full-scale model that are computationally expensive, proposed algorithm is able to obtain solution in less time. For problem 5, proposed approach is able to report feasible solution of 68717 in 2755 seconds while optimal solution of 67286 is obtained by full scale model in 5115 seconds. The solution procedure for problem 5 is shown in Figure 7.5. The solution time for algorithm is driven by relaxed blend scheduling problem (LBSP). More specially, relaxed diesel blending (LBSP1) is the most difficult to solve due to presence of multipurpose blend unit and multipurpose product tanks. Table 7.2 Model statistics for Example 1 Augmented Lagrangian Algorithm Ex. Full Scale Model LBSP model # Event pt. Event pt. n Demand Int./Cont. var. Int./Cont. variables Orders (Constraints) (Constraints) Nonzero Elem. Nonzero Elem. LPSP model Event pt. n Int./Cont. var. (Constraints) Nonzero Elem. 1 4 LBSP1 LBSP2 6 5 6 5 434/3197 127/742 37/243 235/1610 (7748) (2594) (985) (4859) 246 2 3 4 5 6 7 28085 9794 3556 17498 5 5 5 4 360/2663 116/736 30/200 188/1288 (6392) (2545) (752) (3727) 22608 9584 2571 13045 4 4 4 4 286/2129 100/592 26/164 188/1286 (5050) (1967) (587) (3719) 17461 7096 1926 13022 4 4 4 4 301/2217 109/644 32/200 188/1286 (5339) (2134) (689) (3721) 18547 7718 2284 13021 6 6 6 5 484/3453 173/1039 57/347 235/1611 (8653) (3739) (1335) (4862) 31676 14750 4888 17508 6 6 6 5 484/3453 173/1039 57/347 235/1611 (8643) (3729) (1335) (4862) 31586 14661 4888 17508 7 7 7 6 640/4383 266/1511 80/467 282/1936 (11642) (5718) (1930) (6092) 44572 23720 7518 22632 4 4 6 8 8 13 247 Table 7.3 Numerical results for Example 1 Augmented Lagrangian Algorithm Full Scale model (0) (0) Cf(0) 0.2 , CTs 0.3 , CTf 0.2 , 0.37 , 2.15 Ex Objective CPU(sec) Gap Iter. ze (%) m 2 g Time (sec) f e 2 g g 2 1 608.73 501.89 0.00 10 368.02 502.35 11.90 0.70 2 73.29 428.25 0.00 12 525.29 428.19 541.73 0.87 3 16.35 788.98 0.00 9 138.95 791.77 -16.23 0.40 4 36.12 735.16 0.00 15 195.30 747.36 -2.98 0.34 5 5114.95 672.86 0.00 14 2755.25 687.17 1693.37 0.86 6 23085.76 866.43 0.00 8 3001.96 869.75 37.65 0.76 7 28800 695.79 12.91 10 14720.16 695.62 122.12 0.90 248 Figure 7.5 Solution procedure for Example 1, problem 5. Example 2. Refinery configuration with parallel blend unit and no component flow is examined to check efficacy of the proposed decomposition approach. Parameters for model (LPSP) objective function are chosen to be c16 UH e 2 , c18 UH e 1 , and c20 0.75c1j . Parallel diesel 1 1 blenders that can produce three different grade of diesel product are present thus compared to Example 1, this refinery configuration require less number of event points for blend scheduling model (LBSP) to obtain optimal solution that satisfy demand orders for finished blend products. Decomposition provides smaller scheduling problem than full scale scheduling problem as shown in Table 7.4. Proposed decomposition algorithm is able to provide a good feasible solution and in reasonable solution time, as reported in Table 7.5. Solution procedure for problem 5 is shown in 249 Figure 7.6. The solution time in ALO-DQA algorithm is mainly spent solving relaxed problem corresponding to multipurpose diesel blend units. Table 7.4 Model statistics for Example 2 Augmented Lagrangian Algorithm Full Scale Model LBSP model Event pt. Event pt. Int./Cont. var. Int./Cont. variables (Constraints) (Constraints) Nonzero Elem. Nonzero Elem. LPSP model # Ex. Event pt. Orders Int./Cont. var. (Constraints) Nonzero Elem. 1 2 3 4 LBSP1 LBSP2 4 4 4 3 330/2425 122/882 23/159 150/1010 (6319) (3465) (550) (2928) 22522 12906 1788 10218 4 4 4 3 330/2425 122/882 23/159 150/1010 (6319) (3465) (550) (2928) 22522 12906 1788 10218 3 3 3 3 245/1817 106/668 19/124 150/1007 (4655) (2512) (408) (2929) 16274 9059 1280 10205 3 3 3 3 255/1885 112/708 23/152 150/1007 (4883) (2647) (483) (2929) 4 4 4 6 250 5 6 7 17107 9558 1539 10202 5 5 5 4 455/3249 184/1238 46/289 200/1347 (8821) (5084) (1040) (4139) 32386 19672 3636 14870 4 4 4 3 360/2601 140/986 35/231 150/1010 (6953) (3835) (776) (2936) 24884 14237 2598 10230 5 5 5 4 503/3509 224/1454 54/333 200/1347 (9941) (5975) (1223) (4139) 36900 23174 4345 14870 8 8 13 Table 7.5 Numerical results for Example 2 Augmented Lagrangian Algorithm Full Scale model (0) (0) Cf(0) 0.3 , CTs 0.3 , CTf 0.3 , 0.30 , 2.20 Ex Objective CPU(sec) Gap Iter. ze (%) m 2 g Time (sec) f e 2 g g 2 1 559.25 464.88 0.00 8 446.08 474.23 -7.65 0.379 2 82.62 382.35 0.00 10 539.03 413.51 92.67 0.568 3 39.58 629.05 0.00 9 263.35 630.02 -2.50 0.794 4 16.30 579.62 0.00 11 576.03 584.46 -14.43 0.278 5 19466.87 661.53 0.00 16 9700.18 705.76 10580.87 0.971 251 6 12872.25 756.43 0.00 7 2251.59 767.34 -2.49 0.252 7 28800 683.57 13.24 7 13300.39 692.36 -15.76 0.659 Figure 7.6 Solution procedure for Example 2, problem 5 Example 3. Refinery configuration with dedicated component tanks is studied in this example. We include 1 storage tank for each blend component and eliminate all direct streams from production units to blenders. For all problems, parameters value are set at c17 0.01 , c19 0.5 . Model statistics for this example is shown in Table 7.6 and computational results are given in Table 7.7. Augmented Lagrangian algorithm converges to a feasible solution for each problem within 20 iterations and proposed multipliers and penalty parameters gives a solution that is closer to original full-scale model optimal solution. However, for problem 1, the solution obtain using 252 decomposition algorithm is far worse than optimal solution, and this is due to the augmented Lagrangian parameters and penalty. For problem 1, the algorithm converges to a solution that has high task changeovers cost. Convergence profile for problem 5 is given in Figure 7.7. Similar to other example, relaxed diesel blend scheduling (LBSP1) is most difficult to solve and other two problems take only few seconds. Table 7.6 Model statistics for Example 3 Augmented Lagrangian Algorithm Full Scale Model LBSP model Event pt. Event pt. Int./Cont. var. Int./Cont. variables (Constraints) (Constraints) Nonzero Elem. Nonzero Elem. LPSP model # Ex. Event pt. Orders Int./Cont. var. (Constraints) Nonzero Elem. 1 2 3 LBSP1 LBSP2 5 5 5 5 455/2978 169/927 23/159 150/1010 (7230) (3464) (550) (2928) 24793 12582 1788 10218 5 5 4 4 455/2978 122/882 23/159 150/1010 (7230) (3465) (550) (2928) 24793 12906 1788 10218 4 4 4 4 362/2381 106/668 19/124 150/1007 (5715) (2512) (408) (2929) 19193 9059 1280 10205 4 4 4 253 4 5 6 4 4 4 4 377/2469 112/708 23/152 150/1007 (6004) (2647) (483) (2929) 20279 9558 1539 10202 6 6 6 6 598/3831 184/1238 46/289 200/1347 (9664) (5084) (1040) (4139) 34314 19672 3636 14870 6 6 6 5 598/3831 140/986 35/231 150/1010 (9654) (3835) (776) (2936) 34224 14237 2598 10230 6 8 8 Table 7.7 Numerical results for Example 3 Augmented Lagrangian Algorithm Full Scale model (0) (0) (0) KCf 0.05 , KCTs 0.05 , KCTf 0.05 , 0.25 , 2.20 Ex Objective CPU(sec) Gap Iter. ze (%) m 2 g Time (sec) f e 2 g g 2 1 266.31 475.54 0.00 8 1063.97 1743.94 7.96 0.109 2 254.30 421.41 0.00 11 464.76 565.70 -296.52 0.428 3 56.71 730.47 0.00 9 154.27 738.25 0.09 0.202 4 147.78 711.69 0.00 13 239.89 726.78 82.14 0.454 5 28800 654.61 1.06 16 3766.46 658.51 1199.03 0.952 6 28800 866.43 25.18 7 21496.02 871.84 -29.66 0.956 254 Figure 7.7 Solution procedure for Example 3, problem 5. Example 4. In this example, we test the proposed algorithm for a refinery that has blend component tanks and direct flow to blend units. Model statistics and computational results are shown in Table 7.8 and Table 7.9 for problem 5. Table 7.8 Model statistics for Example 5 Full scale Event points Augmented Lagrangian Optimization Approach Model LBSP1 LBSP2 LPSP 6 6 6 5 538 232 57 220 3633 1315 347 1664 Binary variables Continuous 255 Variables Constraints 9097 4863 1335 5021 32774 18599 4888 17791 Nonzero Elements Table 7.9 Numerical results for Example 5 Full scale Model Augmented Lagrangian Optimization Approach CPU time (Sec) 28800 1556 Objective 66895 69938 Gap (%) 10.67 Iterations ( m ) g g 8 2 359.97 g 0.862 Full-scale model is not able to obtain an optimal solution even after 8 hours, whereas proposed approach provides a feasible solution in 1556 seconds. Parameters present in model (LPSP) objective function has value of c16 UH e 2 , c17 0.01 , c18 UH e 1 , 1 Solution procedure for problem 5 is shown in Figure 7.8. 1 and c19 0.5 . 256 Figure 7.8 Solution procedure for Example 4, problem 5. From above results for all four examples, it can be observed that the augmented Lagrangian algorithm converges to a feasible solution of the full-scale problem because the linking variables error norm value always convergences to zero as augmented penalty term . Quality of the feasible solution obtain and solution time is not only affected by augmented Lagrangian optimization parameters but also by parameters c16 , c17 , c18 , c19 , and c20 in some extant. 7.7. Summary We have applied Augmented Lagrangian approach to refinery operations scheduling problem by decomposing the problem into production unit operations and finished blend product and delivery operations. Continuous time formulation based on event-points is used in this work and 257 complicating variables and coupling constraints for decomposed problems are expressed in terms of novel transfer event-points. To improve the performance of the algorithm, a preprocessing step and constraints to incorporate information lost during decomposition are proposed. Proposed approach is shown to effectively address four different refinery network configurations. Main advantages of the proposed decomposition approach are: (a) decomposed problems can be solved using different set of event-points n because complicating variables and coupling variables are expressed in terms of transfer event-points t ; (b) when blending operations network can be decomposed further, this method allows for parallel implementation of blend scheduling problems; (c) the convergence of the algorithm to a feasible solution is guaranteed. Furthermore, this approach can be extended to include crude-oil loading and unloading operations so entire refinery operations scheduling problem can be integrated. Nomenclature Indices bn Blending operations sub-network i Tasks j Production units k Storage units n Event-points m Lagrangian algorithm iteration counter o Product order p Properties s States t Transfer event-points Sets 258 BN All decomposed sub-networks belonging to blending operations Ij Tasks which can be performed in unit j I sc / I sp Tasks which can consume/produce material s J All Production units J ads Blend units that receive at least all components directly from production units J BU Blend units All production units and blend units that receive at least one component J cont directly from production unit operations J ds Blend units that receive at least one component directly from production units J PU Production units that belong to production unit operations J sc / J sp Units that consume/produce material s Ji Units which are suitable for performing task i Jh Units that can produce all the same products as some other unit in the refinery Jm Units which are suitable for performing multiple tasks J seq j Units that follow unit j (no storage in between) J kkp / J kpk Units that consume/produce material s stored in tank k JP Units that produce products K Storage units Kh tanks that can store the same products as some other tank in the refinery pk K kp j / Kj Tanks that store material consumed/produce by unit j Km Multipurpose tanks that can store multiple materials Kp Tanks that can store final products Ks Tanks that can store material s 259 N Event-point within the time horizon Os Orders for finished blend product s that is stored in tanks P Product Properties S States SA Group A products, stored in tanks SB Group B products, not stored in tanks Sbc Blend components states S bn States that belong to sub-network bn Sk States that can be stored in tank k Sic / Sip States that can be consumed/produced by task i S cj / S jp States that can be consumed/produced by unit j Parameters conts , s 1 if components s and s are produced by units in PN that are interconnected Do, s , Do, s Demand limit requirement for order o and product s that is stored in tank max_rates Maximum rate of production of material s oto,i 1 if task i is performed when order o of PUSP is processed by BSP. rs Demand of the final product s at the end of the time horizon RBmin / RBmax j j minimum and maximum production rate of a blend unit j RPjmax Maximum rate material is supplied by (PSP) to blend unit j RPsmax Maximum rate blend component s is supplied by (PSP) to blend unit j max Rimin , j / Ri , j Minimum/ maximum rate of material be processed by task i in unit j RLi Minimum run length for task i RU kmin / RU kmax Minimum/maximum rate of product unloading at tank k 260 stos , k Amount of state s that is present at the beginning of the time horizon in k tak Fill draw delay for product tank k UH Available time horizon Vkmax Maximum available storage capacity of storage tank k Vkheel Maximum heel available for storage tank k yos , k 1 if the material s is present at the beginning of the time horizon in k max smin ,i / s ,i Proportion of state s produced/consumed by task i max bsmin , j / bs , j Minimum/maximum proportion of blend component s consumed by blend unit j s Quality index for finished blend product, higher index means better quality Variables Binary Variables a j , j , n Assigns the materials flow to blend unit j from unit j at event-point n wvi , j , n Assignment of task i in unit j at event-point n ins , j , k , n Assigns the material flow of s into storage tank k from unit j at point n Assigns the starting of product flow out of product tank k to satisfy order o at lk ,o, n event-point n outs , k , j , n Assigns the material flow of s out of storage tank k into unit j at point n ys , k , n Denotes that material s is stored in tank k at event-point n 0-1 continuous variables j ,n For unit j, 1 if the unit becomes active for very first time at event-point n k , n For tank k, 1 if the tank becomes active for very first time at event-point n s , s , k , n Continuous 0-1 variable, 1 if material in tank k switchover service from s at 261 event-point n to s‟ at later event-point os , s, k i , i , j , n Continuous 0-1 variable, 1 if material in tank k switchover service from s to s‟ Continuous 0-1 variable, 1 if task at unit j changes from i at event-point n to i’at later event-point. x j , n,t / xks , n,t Cont. 0-1 variable denotes transfer event-point t corresponds to event-point n Positive variables bps ,i , j , n Amount of material s produced task i in unit j at event-point n bcs ,i , j , n Amount of material s undertaking task i in unit j at event-point n Cf s , j ,t / KCf s ,t Amount of blend component s transferred between BN and PN at event-point t dgol Minimum demand quantity give-away term for order o dgou Maximum demand quantity give-away term for order o H time horizon used for production tasks JJf s , j , j , n Flow of state s from unit j to consecutive unit j’ for consumption at point n Kif s , j ,k ,n Flow of material s from unit j to storage tank k event-point n Kof s , k , j , n Flow of material s from storage tank k to unit j at point n Lfo, k , n Flow of final product for order o from storage tank k at event-point n mhs , k ,n Maximum heel give-away term for product tanks rgs Minimum demand quantity give-away term for Group B product s Rif s , k , n Flow of raw material to storage tank k event-point n sts , k , n Amount of state s present in storage tank k at event-point n Amount of state s that is downgraded to state s‟ in storage tank k at event-point std s , s, k , n n Tearlyo Early fulfillment of order o than required 262 Tfi , j , n Time that task i finishes in unit j at event-point n CTf j ,t / KCTf s ,t Finish time for components transfer between BN and PN at time-point t Tlateo Late fulfillment of order o than required Tosk ,o, n Time that material starts to flow from tank k for order o at event-point n Tof k ,o, n Time that material finishes to flow from tank k for order o at event-point n Tsi , j , n Time that task i starts in unit j at event-point n CTs j ,t / KCTss ,t Start time for components transfer at time-point t Tsf j , k , n Time that material finishes to flow from unit j to tank k at event-point n Tsf k , j , n Time that material finishes to flow from tank k to unit j at event-point n Tss j , k , n Time that material starts to flow from unit j to storage tank k Tssk , j , n Time that material starts to flow from tank k to unit j at event-point n Uif s , j ,n Flow of raw material s to production unit j at point n Uofbs , j , j , n Flow of component s from production unit j to blend unit j at point n Uof s , j , n Flow of product material s from unit j at point n UUf j , s ,v , n Amount of state s received by unit j at period v is processed at event n z Objective value of full-scale integrated model 263 Acknowledgement of Prior Works All of the work presented in this dissertation represents original research by the author. However some of the concepts presented in Chapters 2 to Chapter 6 have previously appeared in published works, for which the author of this dissertation is also the first author. The author of this dissertation is also first author on these prior works, which include: - Shah, N. K., & Ierapetritou, M. G. (In Process). Lagrangian Decomposition Approach to Scheduling of a Large-Scale Oil-Refinery Operations. - Shah, N. K., & Ierapetritou, M. G. (In Process). Efficient Heuristic Algorithm Strategy for Short-Term Scheduling of Refinery Operations. - Shah, N. K., & Ierapetritou, M. G. (2012). Integrated production planning and scheduling optimization of multisite, multiproduct process industry. Computers and Chemical Engineering, 37, 214-226. - Shah, N. K., Li, Z., & Ierapetritou, M. G. (2011). Petroleum Refining Operations: Key Issues, Advances, and Opportunities. Industrial & Engineering Chemistry Research, 50, 1161-1170. - Shah, N. K., & Ierapetritou, M. G. (2011). Short-term scheduling of a large-scale oil-refinery operations: Incorporating logistics details. AIChE Journal, 57, 1570-1584. - Shah, N., Saharidis, G. K. D., Jia, Z., & Ierapetritou, M. G. (2009). Centralized-decentralized optimization for refinery scheduling. Computers & Chemical Engineering, 33, 2091-2105. 264 Appendix Appendix Chapter 2 Table A2-1 Example 1: Process data for production sites Production Unit Capacity Suitability Processing Time Heater 100 Heating 1.0 Reactor 1 50 Reactions 1,2,3 2.0,2.0,1.0 Reactor 2 80 Reactions 1,2,3 2.0,2.0,1.0 Sill 200 Separation Site S1 1 for Product 2 2 for IntAB S2 Heater 150 Heating 1.0 Reactor 1 100 Reactions 1,2,3 2.0,2.0,1.0 Reactor 2 80 Reactions 1,2,3 2.0,2.0,1.0 1 for Product 2 Sill 300 Separation 2 for IntAB S3 Heater 100 Heating 1.0 Reactor 1 75 Reactions 1,2,3 2.0,2.0,1.0 Reactor 2 100 Reactions 1,2,3 2.0,2.0,1.0 1 for Product 2 Sill 150 Separation 2 for IntAB Production Storage State Site Initial Amount Capacity Feed A 100000 100000 Feed B 100000 100000 S1 265 S2 S3 Feed C 100000 100000 Hot A 100 0.0 Int AB 200 0.0 Int BC 150 0.0 Impure 200 0.0 Product 1 400 0.0 Product 2 375 0.0 Feed A 100000 100000 Feed B 100000 100000 Feed C 100000 100000 Hot A 130 0.0 Int AB 250 0.0 Int BC 150 0.0 Impure 300 0.0 Product 1 300 0.0 Product 2 425 0.0 Feed A 100000 100000 Feed B 100000 100000 Feed C 100000 100000 Hot A 115 0.0 Int AB 250 0.0 Int BC 150 0.0 Impure 300 0.0 Product 1 350 0.0 Product 2 400 0.0 266 Table A2-2 Production and Inventory cost data Production Sites Heating Fixed Cost Variable Costs 150 1.0 100 0.5 Reactions S1 1,2,3 Separation 150 1.0 Heating 175 1.5 125 0.75 Reactions S2 1,2,3 S3 Separation 175 1.5 Heating 100 0.8 Inventory Cost P1 10 P2 10 P1 15 P2 15 P1 8 P2 8 Table A2-3 Transportation and Backorder cost data Market M1 M2 M3 Transportation Cost S1 5 S2 8 S3 10 S1 10 S2 5 S3 8 S1 8 S2 10 S3 8 Backorder Cost P1 100 P2 100 P1 150 P2 150 P1 75 P2 75 267 Figure A2-1 Demand data for Example 1 268 Appendix Chapter 3 Table A3-1 Demand in Thousand barrels Scenario 1 Final Products Scenario 2 Case 1 Case 2 Case 3 Case 4 Case 1 Case 2 Case 3 Case 4 FCC gas 10 50 35 75 75 150 200 200 Coke 10 35 10 75 75 75 100 150 Carb diesel 10 100 50 100 100 300 350 500 EPA diesel 10 100 50 150 150 250 250 250 10 100 50 150 150 250 300 250 Red dye diesel Table A3-2 Production rates in thousand barrels/hour Unit Task Rmin Rmax 4CU 4CU Normal 0.1 7.292 2CU 2CU Normal 0.1 4.167 Vacuum Tower Vacuum Normal 0.1 5.708 Coker Coker Normal 0.1 2.75 FCC HDS FCCHDS Normal 0.1 3.00 FCC Maxdistillation Mode 0.1 2.708 FCC Maxgasoline Mode 0.1 2.708 269 Diesel Diesel Normal 0.1 2.5 Isomax Isomax Normal 0.1 2.042 Diesel Blemder Carb Diesel 0.1 54.167 Diesel Blender EPA Diesel 0.1 54.167 Diesel Blender RedDye Diesel 0.1 54.167 Table A3-3 Recipe data State Task State Produced ρP ρc Consumed 4 CU Diesel 0.691 Normal Resid 0.309 2 CU Diesel 0.284 Normal Resid 0.716 Vacuum Resid 0.334 Heavy Gasoil 0.333 Light Gasoil 0.333 Coke 0.640 Heavy Gasoil 0.180 Light Gasoil 0.180 FCC HDS 1 Maxdistillation FCC Gas 0.50 Mode LCO 0.50 ANS1 1 SJV Crude 1 Resid 1 Vacuum Resid 1 Heavy Gasoil 1 FCC HDS 1 Vacuum Tower Normal Coker Normal FCC HDS Normal 270 Maxgosoline FCC Gas 0.50 Mode LCO 0.50 HDS Diesel Hydro Diesel FCC HDS 1 1 Diesel 1 1 Light Gasoil 1 HDS Diesel 0.20 Hydro Diesel 0.25 LCO 0.55 HDS Diesel 0.25 Hydro Diesel 0.15 LCO 0.60 HDS Diesel 0.23 Hydro Diesel 0.12 LCO 0.65 Diesel Normal Isomax Normal Carb Diesel Carb Diesel 1 Normal EPA Diesel EPA Diesel 1 Normal Red Dye Diesel Red Dye Diesel 1 Normal Table A3-4 Storage tank capacity data in thousand barrels Storage Tank Material Stored Capacity Initial Amount ANS Feed Tank 1 ANS 750 10 ANS Feed Tank 2 ANS 750 10 Coker Feed Tank Vacuum Resid 500 0 FCC HDS Feed Tank Heavy Gasoil 500 0 Diesel HDS Feed Tank Diesel 500 0 271 Isomax Feed Tank Light Gasoil 500 0 Carb Diesel Tank 1 Carb Diesel 200 0 Carb Diesel Tank 2 Carb Diesel 200 0 Diesel Swing Tank Carb Diesel 200 0 EPA Diesel Tank 1 EPA Diesel 125 0 EPA Diesel Tank 2 EPA Diesel 125 0 RedDye Diesel Tank1 RedDye Diesel 150 0 RedDye Diesel Tank 2 RedDye Diesel 150 0 Figure A3-1 Gantt chart of tank loading schedule for example 1, Centralized system, scenario 1 272 Figure A3-2 Gantt chart of the tank unloading schedule for example 1, Centralized system, scenario 1 Figure A3-3 Gantt chart of tank loading schedule for example 1, Decentralized system, scenario 1 273 Figure A3-4 Gantt chart of the tank unloading schedule for example 1, Decentralized system, scenario 1 Figure A3-5 Gantt chart of the operation schedule for example 1, Centralized system, scenario 2 274 Figure A3-6 Gantt chart of the operation schedule for example 1, Decentralized system, scenario 2 Figure A3-7 Gantt chart of tank loading schedule for example 1, Centralized system, scenario 2 275 Figure A3-8 Gantt chart of tank unloading schedule for example 1, Centralized system, scenario 2 Figure A3-9 Gantt chart of tank loading schedule for example 1, Centralized system, scenario 2 276 Figure A3-10 Gantt chart of tank unloading schedule for example 1, Centralized system, scenario 2 277 Appendix Chapter 6 Table A6-1 Production rates in thousand barrels/hour Unit Task Rmin Rmax Diesel Blender Carb Diesel 0.1 54.167 Diesel Blender EPA Diesel 0.1 54.167 Diesel Blender RedDye Diesel 0.1 54.167 Penex Penex Normal 0.1 0.792 De-C5 De-C5 Normal 0.1 1.042 Alky Unit Alky Normal 0.1 0.875 Jet Blender Jet Normal 0.1 3.125 CCR CCR Normal 0.1 1.75 Isomax Isomax Normal 0.1 2.042 Naphtha Pre-fractionator Fractionator Normal 0.1 4.167 FCC MaxDistillation Mode 0.1 2.708 FCC MaxGasoline Mode 0.1 2.708 Diesel HDS Diesel Normal 0.1 2.5 Naphtha HDS Naphtha Normal 0.1 3.229 FCC HDS FCCHDS Normal 0.1 3.0 Coker Coker Normal 0.1 2.75 Vacuum Tower Vacuum Normal 0.1 5.708 278 Gas Plant Gas Normal 0.1 4.167 2Crude Distillation Unit 2CU Normal 0.1 4.167 4Crude Distillation Unit 4CU Normal 0.1 7.292 Table A6-2 Minimum/maximum recipe data Task Carb Diesel EPA Diesel RedDye Diesel Penex Normal State Produced Carb Diesel EPA Diesel Red Dye Diesel Isomerate sp,i,min / sp,i,max 1/1 1/1 1/1 1/1 Isomax Gasoline 0.5/1 Pentane 0.5/1 Pentane 0.167/0.167 De-C5 Normal Alky Normal NC4 0.167/0.167 Alkylate 0.1/1 State Consumed sc,,min / sc,,max i i FCC LCO 0.01/0.25 HDS Diesel 0.5/0.75 HydrocrackedDiesel 0.05/0.25 FCC LCO 0.05/0.3 HDS Diesel 0.55/0.8 HydrocrackedDiesel 0.01/0.2 FCC LCO 0.05/0.3 HDS Diesel 0.6/0.85 HydrocrackedDiesel 0.01/0.2 Light Naphtha 0.001/1 Light Ends 0.001/1 Butane 0.001/1 Light Gasoline 1/1 Olefins 0.8/1.125 Iso Butanes 0.2/0.225 279 Jet Normal Jet Fuel 1/1 Light Ends 0.06/0.077 Reformate 0.7/1 HydrocrackedDiesel 0.05/1 Isomax Jet 0.05/1 Heavy Naphtha 0.05/1 Light Gasoline 0.05/1 Olefins 0.05/1 Fractionator Heavy Naphtha 0.001/1 Normal Light Naphtha 0.001/1 FCC HCO 0.05/1 MaxDistillation Olefins 0.05/1 Mode FCC LCO 0.05/1 FCC Gas 0.05/1 FCC HCO 0.05/1 Olefins 0.05/1 CCR Normal Isomax Normal MaxGasoline Mode Isomax Jet 0.001/1 Coker Jet 0.001/1 SR Jet 0.001/1 Heavy Naphtha 1/1 Light Gasoil 1/1 Naphtha HDS 1/1 FCC HDS 1/1 FCC HDS 1/1 FCC LCO 0.05/1 FCC Gas 0.05/1 Diesel Normal HDS Diesel 1/1 Diesel 1/1 Naphtha Normal Naphtha HDS 1/1 Naphtha 1/1 FCCHDS Normal FCC HDS 1/1 Heavy Gasoil 1/1 Coke 0.28/0.28 Vacuum Resid 1/1 Heavy Gasoil 0.1/1 Coker Normal 280 Vacuum Normal Gas Normal 2CU Normal 4CU Normal Light Gasoil 0.1/1 Coker Jet 0.1/1 Naphtha 0.1/1 Vacuum Resid 0.1/1 Heavy Gasoil 0.1/1 Light Gasoil 0.1/1 Iso-Butane 0.001/1 Butane 0.001/1 Gases 0.001/1 Resid 0.53/0.53 Diesel 0.21/0.21 SR Jet 0.11/0.11 Naphtha 0.12/0.12 Light Gas 0.03/0.03 Resid 0.47/0.47 Diesel 0.21/0.21 SR Jet 0.145/0.145 Naphtha 0.14/0.14 Light Gas 0.035/0.035 Resid 1/1 Light Gas 1/1 SJV crude 1/1 Crude oil 1/1 Table A6-3 Storage tank capacity data in thousand barrels Storage Tank Material Stored Capacity Initial Amount ANS Feed Tank 1 ANS 750 10 ANS Feed Tank 2 ANS 750 10 281 Coker Feed Tank Vacuum Resid 500 0 FCC HDS Feed Tank Heavy Gasoil 500 0 Diesel HDS Feed Tank Diesel 500 0 Isomax Feed Tank Light Gasoil 500 0 Carb Diesel Tank 1 Carb Diesel 200 0 Carb Diesel Tank 2 Carb Diesel 200 0 Diesel Swing Tank EPA Diesel 200 - Diesel Swing Tank Carb Diesel 200 - EPA Diesel Tank 1 EPA Diesel 125 0 EPA Diesel Tank 2 EPA Diesel 125 0 RedDye Diesel Tank1 RedDye Diesel 150 0 RedDye Diesel Tank 2 RedDye Diesel 150 0 Naphtha HDS Feed Tank Naphtha 500 0 Jet Product Tank 1 Jet Fuel 200 0 Jet Product Tank 2 Jet Fuel 200 0 282 Table A6-4 Group A products demand order data for the case study Orders (Product type, amount(kbbl), delivery window, delivery rate(kbbl/h)) Orders 1 2 3 4 5 Ex1 Ex2 Ex3 Ex4 Ex5 Ex6 Ex7 Ex8 Ex9 Ex10 P3 P3 P2 P3 P1 P3 P1 P3 P1 P2 [60,80] [50,70] [100,120] [90,100] [120,125] [60,75] [25,30] [15,25] [40,50] [20,55] [30,50] [20,35] [15,25] [50,75] [40,75] [28.5,46.5] [12,30] [30,50] [30,40] 20,35] P2 P4 P4 P4 P2 P2 P1 P2 P2 P2 [50,70] [50,90] [50,90] [140,150] [100,110] [90,100] [50,65] [40,50] [40,55] [35,50] [60,90] [45,80] [45,80] [60,90] [90,115] [65,89] [50,70] [45,75] [45,58] [40,55] P4 P2 P3 P1 P1 P4 P2 P4 P1 P1 [50,80] [50,75] [50,75] [100,115] [150,175] [140,150] [45,50] [50,75] [45,53] [40,50] [75,100] [60,80] [60,80] [100,120] [135,150] [75,91] [90,105] [75,110] [70,85] [60,75] P1 P1 P1 P4 P2 P1 P4 P2 P3 P3 [50,75] [50,80] [50,80] [100,120] [90,120] [90,115] [90,110] [20,30] [40,44] [15,60] [75,110] [70,90] [70,90] [125,150] [168,185] [102,116] [98,128] [75,95] [92,105] [80,100] P2 P3 P4 P3 P3 P2 P2 [110,125] [105,125] [100,150] [50,65] [10,20] [30,60] [30,50] [130,150] [200,220] [105,130] [125,141] [90,125] [110,125] [105,115] 283 6 7 8 9 10 11 P1 P4 P1 P2 P1 P3 P3 [50,60] [120,150] [100,115] [35,40] [15,30] [35,50] [40,45] [160,175] [50,75] [125,153] [160,180] [100,115] [120,140] [110,125] P4 P4 P4 P1 P1 P1 [100,120] [100,165] [100,115] [55,60] [35,40] [30,65] [120,146] [155,170] [160,193] [115,145] [145,160] [130,145] P4 P2 P3 P4 P3 P1 [160,170] [75,100] [45,65] [75,100] [30,75] [25,50] [175,205] [179,203] [175,205] [115,145] [175,190] [150,165] P4 P4 P4 P2 P2 [120,125] [85,100] [80,100] [40,40] [30,45] [195,210] [200,219] [150,175] [195,210] [170,185] P3 P1 P2 P1 P3 [75,100] [70,90] [50,65] [55,75] [25,55] [200,215] [210,227] [155,170] [205,220] [170,190] P2 P3 P2 P1 [50,65] [10,15] [40,45] [20,55] [220,240] [160,180] [205,225] [200,220] 284 12 13 14 15 16 17 P4 P3 P1 P1 [120,140] [35,40] [50,60] [25,45] [223,240] [180,200] [220,240] [215,235] P3 P2 P4 P3 [60,80] [10,15] [50,75] [30,40] [228,240] [200,220] [50,75] [220,240] P4 P4 P4 [100,115] [75,115] [40,60] [200,230] [100,120] [30,65] P1 P4 P4 [30,45] [60,110] [55,95] [210,235] [125,150] [55,75] P4 P4 P4 [75,100] [55,120] [50,65] [220,240] [140,165] [95,110] P4 P4 [85,125] [30,105] [175,200] [115,125] 285 18 19 P4 P4 [100,125] [60,125] [205,225] [130,155] P4 P4 [78,100] [75,105] [220,240] [160,180] P4 20 [85,115] [180,205] P4 21 [100,120] [200,225] P4 22 [110,130] [215,240] Table A6-5 Group B products demands data for the case study Ex. Initial holdup in swing Group B products demand (kbbl) 286 tank (product, kbbl) P5 P6 P7 P8 P9 P10 P11 P12 P13 1 - 5 5 10 10 10 5 4 5 5 2 P1 - 10 10 5 10 15 10 10 5 15 10 3 P1 - 10 5 10 15 10 5 5 10 25 5 4 - 10 25 25 20 5 10 10 15 0 5 - 15 30 25 25 15 13 10 50 10 6 - 13 25 30 10 20 15 17 75 5 7 P1 - 10 10 20 30 25 30 15 10 60 15 8 - 20 15 25 20 25 10 15 95 20 9 P1 - 10 15 35 35 20 20 15 20 70 10 10 - 10 25 35 25 15 16 10 80 10 287 Appendix Chapter 7 Valid Inequalities There are several constraints, redundant in the model but including them improves computational time. We propose following inequalities for decomposed models: x j , n ,t t t , n t x j , n,t , j J BU J ds , n N , t T , n t , t 1, n 1 x j , n,t x j , n,t , j J BU J ds , n N , t T , t T , t t , n t (1b) n t xks , n ,t xks , n,t , s Sk , n N , t T , n t , t 1, n 1 t t , nt (1a) (1c) xks , n,t xks , n,t , s Sk , n N , t T , t T , t t , n t (1d) n t Eqs. (1a)-(1d) restates that transfer event-point t t should be assigned an active event-point n before t and n n should be assigned a transfer event-point before n . These inequalities are added to both (LBSP) and (LPSP). Relaxed Blend Scheduling Problem: ys , k , n in s, j,k ,n yos , k in s, j,k ,n ys , k , n 1 jJ kpk ys , k , n jJ kpk s Sk ,s s yos, k , s S A S kK m Sbc , k K s , n 1 s Sk ,s s ys, k , n 1 , s S A S kK m Sbc , k K s , n N , n 1 (2a) (2b) Inequalities (2a)-(2b) states that if material is stored in tank at event point n then it either be hold up from previous event-point or flow of the material in to the tank at the same event-point. Above constraint is only included for blend components and finished blend products that can be stored in multipurpose (swing) product tanks ( K m ). wvi , j , n T 1 wvi , j , n , bn BN , n 1, j J ds jJ ds jJ ds ,iI j (3a) wvi , j , n N 1 wvi , j , n , bn BN , n 1, j J ds jJ BU , jJ ds jJ ds , iI j (3b) jJ BU J ds ,iI j , n 1 jJ BU / J ds , iI j , n 1 288 wv T , j J BU J ds , xsP, j ,t T i, j ,n iI j , n (3c) s in T , s Sbc , k K s , j J dmy , xk s ,t T P s, j ,k ,n n (3d) s We restrict a feasible solution of decomposed scheduling problem by always requiring the production to take place during first event-point using Eqs. (3a) and (3b) and these inequalities do not eliminate optimal solution. For blend unit that receive at least one blend components straight from production unit operations without any storage, the blend unit would be operational only when it is receiving component from PN. Thus, the total run-modes for the blend unit are bounded by total transfer event-points as stated by Eq. (3c). Similarly, inequality (3d) stresses the flow of component into a tank also bounded. x nt B j , n ,t xk n t 1, j J BU J ds , t T B s , n ,t 1, s S k , t T x B j , n ,t x B j , n ,t wvi , j , n , j J BU J ds , n N n t t n iI j , n t iI j xk B s , n ,t xk B s , n ,t n t t n k K s , jJ dmy , n t xk k K s t n wvi , j , n , j J BU J ds , t T k K s , jJ dmy B s , n ,t ins , j .k , n , s S k , t T ins , j .k , n , s Sk , n N k K s , jJ dmy ins , j .k , n , s S k , n N (4a) (4b) (4c) (4d) (4e) (4f) (4g) If flow is happening between PN and BN at event-point n then it must corresponds to at most one transfer event-point t as restated by Eqs. (4a) and (4b). If there is flow at event point t then it must correspond to some event-point n and this is reinforced by Eqs. (4c)-(4g). Relaxed Production Unit Scheduling Problem: 289 bn sSbc ,iI sp , j J sc J ds , jJ i , n 1 wvi , j , n T 1 wvi , j , n , bn bn sSbc , jJ sp , j J sc J ds sSbc ,iI sp , j J sc , jJ i bn BN , n 1, j J ds (5) Valid inequality (5) limits the feasible solution space by requiring production to take place during first event-point ( n 1 ) for production units that supply components directly to a blend unit. a j , j , n a j , j , n wv i, j,n j J BU J sc j J BU , s Sbc / Sk , j J sp , n N j J BU J sc iI sp J sc wv i, j ,n , s Sbc / S k , j J sp , n N iI sp (6a) (6b) If a blend component that has no storage option is being produced by a production unit, then the production unit will be sending this blend component to suitable blend unit and this material balance relation is restated in valid inequalities (6a) and (6b). x t n P j , n ,t wvi , j , n , s Sbc / S k , j J BU , n N iI sp , j J i x Pj , n ,t j J sp , t n nN , t n wvi , j , n , s Sbc / Sk , j J BU / J h , n N (7b) wvi , j , n , s Sbc / Sk , j J BU / J h (7c) (7d) iI sp , x Pj , n ,t iI sp , P j , n ,t x P j , n ,t t n t n j J i j J i , n x Pj , n ,t BU jJ bn , j J sp ,t n x wvi , j , n , s Sbc / Sk , bn BN , n N , j J bnBU J h iI sp , j J i a j , j , n , s Sbc / Sk , j J BU , n N j J sp xk t n a j , j , n , s Sbc / Sk , j J BU , j J sp , n N P s , n ,t xk sP, n ,t xksP, n ,t nN , t n wvi , j , n , s Sbc S k , n N iI sp , jJ i jJ sp , t n (7a) (7e) (7f) (7g) wvi , j , n , s Sbc Sk , n N (7h) wvi , j , n , s Sbc Sk (7i) iI sp , jJ i iI sp , jJ i , n Valid inequalities (7a)-(7f) restates the relationship between production units that supply components directly to blend units and transfer event-point t and production event-point n . Eqs. 290 (7g)-(7i) emphasis relationship between production units that supply components to BN for storage at event-point n and to transfer event-point t . These inequalities speed up convergence to optimal solution by decreasing the number of iterations and nodes in branch and bound tree in solver CPLEX. When the optimization model is written in GAMS environment, where the variables and equations are declared have significant impact on the solution time (GAMS TUTIORIAL). In our work, we have added all the inequalities presented in this section to our relaxed models (LBSP) and (LPSP). However, depending upon how the model is written in GAMS, one or more of these inequalities might not be necessary. Table A7-1 Production rates (thousand barrels/hour) for a second diesel blend unit Diesel Blender2 Carb Diesel 0.1 34.613 Diesel Blender2 EPA Diesel 0.1 34.613 Diesel Blender2 RedDye Diesel 0.1 34.613 Table A7-2 Storage tank capacity data for blend component tanks Storage Tank Material Stored Capacity Initial Amount FCC LCO Tank FCC LCO 300 0 HDS Diesel Tank HDS Diesel 500 0 HydrocrackedDiesel Tank HydrocrackedDiesel 350 0 Isomax Jet Tank Isomax Jet 500 0 Coker Jet Tank Coker Jet 500 0 SR Jet Tank SR Jet 500 0 291 Table A7-3 Group B products demands data for the case study Group B products demand (kbbl) Initial holdup Ex. in swing tank P5 P6 P7 P8 P9 P10 P11 P12 P13 (product, kbbl) 1 - 5 5 10 10 10 5 4 5 5 2 - 2 14 30 2 4 3 6 4 30 3 P1 - 10 10 5 10 15 10 10 5 15 10 4 - 10 25 25 20 5 10 10 15 0 5 - 15 30 25 25 15 13 10 50 10 6 P1-10 12 25 25 15 15 12 10 45 10 7 P1 - 10 10 20 30 25 30 15 10 60 15 292 Table A7-4 Group A products demand order data for the case study Orders (Product type, amount(kbbl), delivery window, delivery rate(kbbl/h)) Orders 1 2 3 4 5 Ex1 Ex2 Ex3 Ex4 Ex5 Ex6 Ex7 P3 P3 P3 P3 P1 P3 P1 [60,80] [60,80] [50,70] [90,100] [120,125] [100,115] [25,30] [30,50] [30,50] [20,35] [50,75] [40,75] [40,70] [12,30] P2 P2 P4 P4 P2 P2 P1 [50,70] [50,70] [50,90] [140,150] [100,110] [100,110] [50,65] [60,90] [60,90] [45,80] [60,90] [90,115] [70,100] [50,70] P4 P4 P2 P1 P1 P1 P2 [50,80] [50,80] [50,75] [100,115] [150,175] [150,175] [45,50] [75,100] [75,100] [60,80] [100,120] [135,150] [115,145] [90,105] P1 P1 P1 P4 P2 P2 P4 [50,75] [50,75] [50,80] [100,120] [90,120] [90,120] [90,110] [75,110] [75,110] [70,90] [125,150] [168,185] [180,215] [98,128] P2 P3 P1 P3 [110,125] [105,125] [100,115] [50,65] [130,150] [200,220] [200,225] [125,141] 293 6 7 8 P1 P4 P4 P2 [50,60] [120,150] [110,130] [35,40] [160,175] [50,75] [60,98] [160,180] P4 P4 P4 [100,120] [120,140] [100,115] [120,146] [124,155] [160,193] P4 P4 P3 [160,170] [160,170] [45,65] [175,205] [195,235] [175,205] P4 9 [85,100] [200,219] P1 10 [70,90] [210,227] P2 11 [50,65] [220,240] 294 P4 12 [120,140] [223,240] P3 13 [60,80] [228,240] 295 Bibliography ACC. (2014). Chemistry industry fact. In. http://www.americanchemistry.com/chemistryindustry-facts: American Chemistry Council. Adhya, N., Tawarmalani, M., & Sahinidis, N. V. (1999). A Lagrangian Approach to the Pooling Problem. Industrial & Engineering Chemistry Research, 38, 1956-1972. Amaldi, E., Bruglieri, M., & Casale, G. (2008). A two-phase relaxation-based heuristic for the maximum feasible subsystem problem. Computers & Operations Research, 35, 14651482. Andreani, R., Birgin, E. G., Mart´ınez, J. M., & Schuverdt, M. L. (2008). On Augmented Lagrangian methods with general lower-level constraints. SIAM Journal on Optimization, 18, 1286-1309. Arief, A., Rajagopalan, S., & Karimi, I. A. (2007). Heuristic rescheduling of crude oil operations to manage abnormal supply chain events. AIChE Journal, 53, 397-422. Audet, C., Brimberg, J., Hansen, P., Digabel, S. L., & Mladenovic, N. (2004). Pooling Problem: Alternate Formulations and Solution Methods. Management Science, 50, 761-776. Bassett, M. H., Pekny, J., & Reklaitis, G. (1996). Decomposition Techniques for the Solution of Large-Scale Scheduling Problems. American Institute of Chemical Engineers Journal, 42, 3373-3387. Benders, J. F. (1962). Partitioning Procedures for Solving Mixed-Variables Programming Problems. Numerische Mathematic, 4, 238-252. Bertsekas, D. P. (1995). Nonlinear Programming (2nd ed. ed.). Belmont, Massachusetts: Athena Scientific. Bertsekas, D. P., & Tsitsiklis, J. N. (1989). Parallel and Distributed Computation. Englewood Cliffs, New Jersey: Prentice-Hall. Bertsimas, D., & Sim, M. (2003). Robust Discrete optimization and Network Flows. Math. Prog., 98, 49-71. Burkard, R. E., Fortuna, T., & Hurkens, C. A. J. (2002). Makespan minimization for chemical batch processes using non-uniform time grids. Computers & Chemical Engineering, 26, 1321-1332. Calfa, B. A., Agarwal, A., Grossmann, I. E., & Wassick, J. M. (2013). Hybrid BilevelLagrangean Decomposition Scheme for the Integration of Planning and Scheduling of a Network of Batch Plants. Industrial & Engineering Chemistry Research, 52, 2152-2167. Castillo, P. A. C., & Mahalec, V. (2014a). Inventory pinch based, multiscale models for integrated planning and scheduling-part I: Gasoline blend planning. AIChE Journal, 60, 2158-2178. Castillo, P. A. C., & Mahalec, V. (2014b). Inventory pinch based, multiscale models for integrated planning and scheduling-part II: Gasoline blend scheduling. AIChE Journal, 60, 2475-2497. Castro, P. M., & Grossmann, I. E. (2014). Global Optimal Scheduling of Crude Oil Blending Operations with RTN Continuous-time and Multiparametric Disaggregation. Industrial & Engineering Chemistry Research, 53, 15127-15145. Castro, P. M., Harjunkoski, I., & Grossmann, I. E. (2009). Optimal Short-Term Scheduling of Large-Scale Multistage Batch Plants. Industrial & Engineering Chemistry Research. Castro, P. M., Harjunkoski, I., & Grossmann, I. E. (2011). Greedy algorithm for scheduling batch plants with sequence-dependent changeovers. AIChE Journal, 57, 373-387. Chen, J. M., & Chen, T. H. (2005). The Multi-item Replenishment Problem in a Two-Echelon Supply Chain: The Effect of Centralization versus Decentralization. Computers & Operations Research, 32, 3191-3207. 296 Cornillier, F., Boctor, F. F., Laporte, G., & Renaud, J. (2008). A heuristic for the multiperiodpetrol station replenishment problem. European Journal of Operational Research, 191, 295-305. Cornuejols, G., Fisher, M., & Nemhauser, G. L. (1977). On the Uncapacitated Location Problem*. In E. L. J. B. H. K. P.L. Hammer & G. L. Nemhauser (Eds.), Annals of Discrete Mathematics (Vol. Volume 1, pp. 163-177): Elsevier. Cuiwen, C., Xingsheng, G., & Zhong, X. (2013). A data-driven rolling-horizon online scheduling model for diesel production of a real-world refinery. AIChE Journal, 59, 1160-1174. Dimitriadis, A. D., Shah, N., & Pantelides, C. C. (1997). RTN-based rolling horizon algorithms for medium term scheduling of multipurpose plants. Computers & Chemical Engineering, 21, S1061-S1066. Fisher, M. L. (1973). Optimal Solution of Scheduling Problems Using Lagrange Multipliers: Part I. Operations Research, 21, 1114-1127. Fisher, M. L. (1981). The Lagrangian Relaxation Method for Solving Integer Programming Problems. Management Science, 27, 1-18. Fisher, M. L. (1985). An Applications Oriented Guide to Lagrangian Relaxation. Interfaces, 15, 10-21. Floudas, C. A., & Lin, X. (2004). Continuous-Time Versus Discrete-Time Approaches For Scheduling of Chemical Processes: a Review. Computers & Chemical Engineering, 28, 2109-2129. Fortin, M., & Glowinski, R. (1983). Augmented Lagrangian Methods: Application to the Numerical Solution of Boundary-Value Problems: North-Holland. Gaivoronski, A. (1988). Stochastic quasigradient methods and their implementation. In Numerical techniques for stochastic optimization (Vol. 10, pp. 313-351): Springer. Gary, J. H., Handwerk, G. E., & Kaiser, M. J. (2007). Petroleum Refining: Technology and Economics: CRC Press; 5 edition. Ghaddar, B., Naoum-Sawaya, J., Kishimoto, A., Taheri, N., & Eck, B. (2014). Innovative Applications of O.R.: A Lagrangian decomposition approach for the pump scheduling problem in water networks. European Journal of Operational Research. Glismann, K., & Gruhn, G. (2001). Short-term scheduling and recipe optimization of blending processes. Computers & Chemical Engineering, 25, 627-634. Gnoni, M. G., Iavagnilio, R., Mossa, G., Mummolo, G., & Di Leva, A. (2003). Production planning of a multi-site manufacturing system by hybrid modelling: A case study from the automotive industry. International Journal of Production Economics, 85, 251-262. Gounaris, C. E., Misener, R., & Floudas, C. A. (2009). Computational Comparison of Piecewise Linear Relaxations for Pooling Problems. Industrial & Engineering Chemistry Research, 48, 5742-5766. Graves, S. C. (1982). Using Lagrangean Techniques to Solve Hierarchical Production Planning Problems. Management Science, 28, 260-275. Grossmann, I. (2005). Enterprise-wide Optimization: A New Frontier in Process Systems Engineering. AIChE Journal, 51, 1846-1857. Grossmann, I. E. (2014). Challenges in the application of mathematical programming in the enterprise-wide optimization of process industries. Theoretical Foundations of Chemical Engineering, 48, 555-573. Gupta, A., & Maranas, C. D. (1999). A Hierarchical Lagrangean Relaxation Procedure for Solving Midterm Planning Problems. Industrial & Engineering Chemistry Research, 38, 1937-1947. Gupta, A., Maranas, C. D., Rafiqul, G., & Sten Bay, J. (2001). Multiperiod planning of multisite supply chains under demand uncertainty. In COMPUTER AIDED CHEMICAL ENGINEERING (Vol. Volume 9, pp. 871-876): Elsevier. 297 Harjunkoski, I., & Grossmann, I. E. (2001). A Decomposition Approach For the Scheduling of a Steel Plant Production. Computers and Chemical Engineering, 25, 1647-1660. Harjunkoski, I., Maravelias, C. T., Bongers, P., Castro, P. M., Engell, S., Grossmann, I. E., Hooker, J., Méndez, C., Sand, G., & Wassick, J. (2014). Scope for industrial applications of production scheduling models and solution methods. Computers & Chemical Engineering, 62, 161-193. He, Y., & Hui, C.-W. (2007). Genetic algorithm for large-size multi-stage batch plant scheduling. Chemical Engineering Science, 62, 1504-1523. Held, M., & Karp, R. (1971). The traveling-salesman problem and minimum spanning trees: Part II. Mathematical Programming, 1, 6-25. Held, M., & Karp, R. M. (1970). The Traveling-Salesman Problem and Minimum Spanning Trees. Operations Research, 18, 1138-1162. Holmberg, K., & Hellstrand, J. (1998). Solving the Uncapacitated Network Design Problem by a Lagrangean Heuristic and Branch-and-Bound. Operations Research, 46, 247-259. Holmberg, K., & Yuan, D. (2000). A Lagrangian Heuristic Based Branch-and-Bound Approach for the Capacitated Network Design Problem. Operations Research, 48, 461-481. Ierapetritou, M. G., & Floudas, C. (1998 ). Effective Continuous-time Formulation for ShortTerm Scheduling: 1 Multipurpose Batch Processes Industrial & Engineering Chemical Research, 37, 4341-4359. Ierapetritou, M. G., & Floudas, C. A. (1998). Effective Continuous-Time Formulation for ShortTerm Scheduling. 2. Continuous and Semicontinuous Processes. Industrial & Engineering Chemistry Research, 37, 4360-4374. Ierapetritou, M. G., & Floudas, C. A. (1998). Effective Continuous-Time Formulation for ShortTerm Scheduling: 2 Multipurpose/Multiproduct Continuous Processes. Industrial & Engineering Chemical Research, 37, 4360-4374. Ierapetritou, M. G., Hene, T. S., & Floudas, C. A. (1999). Effective Continuous-Time Formulation for Short -Term Scheduling: 3. Multiple Intermediate Due Dates. Industrial & Engineering Chemical Research, 38, 3446-3461. Jackson, J. R., & Grossmann, I. E. (2003). Temporal Decomposition Scheme for Nonlinear Multisite Production Planning and Distribution Models. Industrial & Engineering Chemistry Research, 42, 3045-3055. Janak, S. L., Floudas, C. A., Kallrath, J., & Vormbrock, N. (2006). Production Scheduling of a Large-Scale Industrial Batch Plant. I. Short-Term and Medium-Term Scheduling. Industrial & Engineering Chemistry Research, 45, 8234-8252. Janak, S. L., Lin, X., & Floudas, C. A. (2004). Enhanced Continuous-Time Unit-Specific EventBased Formulation for Short-Term Scheduling of Multipurpose Batch Processes: Resource Constraints and Mixed Storage Policies. Industrial & Engineering Chemistry Research, 43, 2516-2533. Jia, Z., & Ierapetritou, M. (2003). Mixed-Integer Linear Programming Model for Gasoline Blending and Distribution Scheduling. Industrial & Engineering Chemistry Research, 42, 825-835. Jia, Z., & Ierapetritou, M. (2004). Efficient short-term scheduling of refinery operations based on a continuous time formulation. Computers & Chemical Engineering, 28, 1001-1019. Jia, Z., Ierapetritou, M., & Kelly, J. D. (2003). Refinery Short-Term Scheduling Using Continuous Time Formulation: Crude-Oil Operations. Industrial & Engineering Chemistry Research, 42, 3085-3097. Joly, M., Moro, L. F. L., & Pinto, J. M. (2002). Planning and scheduling for petroleum refineries using mathematical programming. Brazilian Journal of Chemical Engineering, 19, 207228. Kallrath, J. (2002a). Combined strategic and operational planning – an MILP success story in chemical industry. OR Spectrum, 24, 315-341. 298 Kallrath, J. (2002b). Planning and scheduling in the process industry. OR Spectrum, 24, 219-250. Karuppiah, R., Furman, K. C., & Grossmann, I. E. (2008). Global optimization for scheduling refinery crude oil operations. Computers & Chemical Engineering, 32, 2745-2766. Karuppiah, R., Martín, M., & Grossmann, I. E. (2010). A simple heuristic for reducing the number of scenarios in two-stage stochastic programming. Computers & Chemical Engineering, 34, 1246-1255. Kelly, J. D. (2003a). Next-generation refinery scheduling technology. NPRA Plant Automation and Decision Support Conference, San Antonio, TX, Presentation. Kelly, J. D. (2003b). Smooth-and-dive accelerator, a pre-MILP primal heuristic applied to scheduling. Computer & Chemical Engineering 27, 873-832. Kelly, J. D. (2006). Logistics: the missing link in blend scheduling optimization. Hydrocarbon Processing, 45-51. Kelly, J. D., & Mann, J. L. (2003a). Crude-oil blend scheduling optimization: An application with multi-million dollar benefits: Part I. Hydrocarbon Processing, 47-53. Kelly, J. D., & Mann, J. L. (2003b). Crude-oil blend scheduling optimization: An application with multi-million dollar benefits: Part II. Hydrocarbon Processing, 72-79. Kelly, J. D., & Zyngier, D. (2008). Hierarchical Decomposition Heuristic for Scheduling: Coordinated Reasoning for Decentralized And Distributed Decision-Making Problems. Computers and Chemical Engineering, 32, 2684-2705. Khor, C. S., Chachuat, B., & Shah, N. (2012). A superstructure optimization approach for water network synthesis with membrane separation-based regenerators. Computers & Chemical Engineering, 42, 48-63. Kolodziej, S. P., Grossmann, I. E., Furman, K. C., & Sawaya, N. W. (2013). A discretizationbased approach for the optimization of the multiperiod blend scheduling problem. Computers & Chemical Engineering, 53, 122-142. Kondili, E. (1987). Optimal Scheduling of Batch Processes. Imperial College London, U.K. Kondili, E., Pantelides, C. C., & Sargent, R. W. H. (1993). A General Algorithm for Short-Term Scheduling of Batch Operations--I. MILP Formulation. Computers & Chemical Engineering, 17, 211-227. Kreipl, S., & Pinedo, M. (2004). Planning and Scheduling in Supply Chains: An Overview of Issues in Practice. Production and Operations Management, 13, 77-92. Lee, H. M., Pinto, J. M., Grossmann, I. E., & Park, S. (1996). Mixed-integer linear programming model for refinery short-term scheduling of crude oil unloading with inventory management. Industrial & Engineering Chemistry Research, 35, 1630-1641. Li, J., & Karimi, I. A. (2011). Scheduling Gasoline Blending Operations from Recipe Determination to Shipping Using Unit Slots. Industrial & Engineering Chemistry Research, 50, 9156-9174. Li, J., Karimi, I. A., & Srinivasan, R. (2010). Recipe determination and scheduling of gasoline blending operations. AIChE Journal, 56, 441-465. Li, Y., Lu, Z., & Michalek, J. J. (2008). Diagonal Quadratic Approximation for Parallelization of Analytical Target Cascading. Journal of Mechanical Design, 130, 051402-051411. Li, Z., & Ierapetritou, M. G. (2010a). Production planning and scheduling integration through augmented Lagrangian optimization. Computers & Chemical Engineering, 34, 996-1006. Li, Z., & Ierapetritou, M. G. (2010b). Rolling horizon based planning and scheduling integration with production capacity consideration. Chemical Engineering Science, 65, 5887-5900. Lin, J., & Chen, Y.-Y. (2007). A multi-site supply network planning problem considering variable time buckets– A TFT-LCD industry case. The International Journal of Advanced Manufacturing Technology, 33, 1031-1044. Luh, P. B., & Hoitomt, D. J. (1993). Scheduling of manufacturing systems using the Lagrangian relaxation technique. Automatic Control, IEEE Transactions on, 38, 1066-1079. 299 Luo, C., & Rong, G. Hierarchical Approach for Short-Term Scheduling in Refineries. Industrial & Engineering Chemistry Research, 49, 4474-4474. Luo, C., & Rong, G. (2007). Hierarchical Approach for Short-Term Scheduling in Refineries. Industrial & Engineering Chemistry Research, 46, 3656-3668. Maravelias, C. T., & Grossmann, I. E. (2003). New General Continuous-Time State鈭扵ask Network Formulation for Short-Term Scheduling of Multipurpose Batch Plants. Industrial & Engineering Chemistry Research, 42, 3056-3074. Maravelias, C. T., & Sung, C. (2009). Integration of production planning and scheduling: Overview, challenges and opportunities. Computers & Chemical Engineering, 33, 19191930. Mendez, C. A., Grossmann, I. E., Harjunkoski, I., & Kabore, P. (2006). A simultaneous optimization approach for off-line blending and scheduling of oil-refinery operations. Computers & Chemical Engineering, 30, 614-634. Minoux, M. (1986). Mathematical Programming Theory and Algorithms. Moniz, S., Barbosa-Póvoa, A. P., De Sousa, J. P., & Duarte, P. (2014). Solution methodology for scheduling problems in batch plants. Industrial and Engineering Chemistry Research, 53, 19265-19281. Moro, L. F. L., Zanin, A. C., & Pinto, J. M. (1998). A planning model for refinery diesel production. Computers & Chemical Engineering, 22, S1039-S1042. Mouret, S., Grossmann, I. E., & Pestiaux, P. (2010). Time Representations and Mathematical Models for Process Scheduling Problems. Computers & Chemical Engineering, 35, 10381063. Mouret, S., Grossmann, I. E., & Pestiaux, P. (2011). A new Lagrangian decomposition approach applied to the integration of refinery planning and crude-oil scheduling. Computers & Chemical Engineering, 35, 2750-2766. Neiro, S. M. S., Murata, V. V., & Pinto, J. M. (2014). Hybrid Time Formulation for Diesel Blending and Distribution Scheduling. Industrial & Engineering Chemistry Research, 53, 17124-17134. Nishi, T., Shinozaki, R., & Konishi, M. (2008). An augmented Lagrangian approach for distributed supply chain planning for multiple companies. Ieee Transactions on Automation Science and Engineering, 5, 259-274. . The Outlook for Energy: A View to 2040. In. (2013): ExxonMobil. Pantelides, C. C. (1994). Unified Frameworks for Optimal Process Planning and Scheduling. In D. W. T. Rippin & J. Hale (Eds.), Proceedings of the Second Conference on Foundations of Computer-Aided Process Operations (pp. 253): CACHE Publications. Papageorgiou, L. G. (2009). Supply chain optimisation for the process industries: Advances and opportunities. Computers & Chemical Engineering, 33, 1931-1938. Pinto, J. M., Joly, M., & Moro, L. F. L. (2000). Planning and scheduling models for refinery operations. Computers & Chemical Engineering, 24, 2259-2276. Polasek, M., & Mann, J. (2013). Distillate blending optimization. In. White Paper: Honeywell Process Solutions. Reddy, P. C. P., Karimi, I. A., & Srinivasan, R. (2004a). A novel solution approach for optimizing crude oil operations. A.I.Ch.E. Journal 50, 1177-1197. Reddy, P. C. P., Karimi, I. A., & Srinivasan, R. (2004b). A new continuous-time formulation for scheduling crude oil operations. Chemical Engineering Science, 59, 1325-1341. Roslof, J., Harjunkoski, I., Westerlund, T., & Isaksson, J. (2002). Solving a large-scale industrial scheduling problem using MILP combined with a heuristic procedure. European Journal of Operational Research, 138, 29-42. Ruszczynski, A. (1995). On convergence of an augmented Lagrangian decomposition method for sparse convex optimization. Mathematics of Operations Research, 20, 634-656. 300 Saharidis, G. K., Dallery, Y., & Karaesmen, F. (2006). Centralized versus Decentralized Production Planning. RAIRO Operations Research, 40. Saharidis, G. K. D., & Ierapetritou, M. G. (2009). Scheduling of Loading and Unloading of Crude Oil in a Refinery with Optimal Mixture Preparation. Industrial & Engineering Chemistry Research, 48, 2624-2633. Saharidis, G. K. D., Kouikoglou, V. S., & Dallery, Y. (2009). Centralized and decentralized control polices for a two-stage stochastic supply chain with subcontracting. International Journal of Production Economics, 117, 117-126. Saharidis, G. K. D., Minoux, M., & Dallery, Y. (2009). Scheduling of loading and unloading of crude oil in a refinery using event-based discrete time formulation. Computers & Chemical Engineering, 33, 1413-1426. Saharidis, G. K. D., Minoux, M., & Ierapetritou, M. G. (2010). Accelerating Benders method using covering cut bundle generation. International Transactions in Operational Research, 17, 221-237. Shah, N. (1998). Single and multisite planning and scheduling: current status and future challenges. AIChE Symposium Series, 94, 75. Shah, N. (2004). Pharmaceutical supply chains: key issues and strategies for optimisation. Computers & Chemical Engineering, 28, 929-941. Shah, N. (2005). Process industry supply chains: Advances and challenges. Computers & Chemical Engineering, 29, 1225-1236. Shah, N. K., Li, Z., & Ierapetritou, M. G. (2011). Petroleum Refining Operations: Key Issues, Advances, and Opportunities. Industrial & Engineering Chemistry Research, 50, 11611170. Shaik, M. A., & Floudas, C. A. (2007). Improved Unit-Specific Event-Based Continuous-Time Model for Short-Term Scheduling of Continuous Processes: Rigorous Treatment of Storage Requirements. Industrial & Engineering Chemistry Research, 46, 1764-1779. Shaik, M. A., Floudas, C. A., Kallrath, J., & Pitz, H.-J. (2009). Production Scheduling of a LargeScale Industrial Continuous Plant: Short-Term and Medium-Term Scheduling. Computers & Chemical Engineering, 33, 670-686. Shapiro, J. F. (2001). Modeling the Supply Chain. Duxbury: Pacific Grove. Stefansson, H., Sigmarsdottir, S., Jensson, P., & Shah, N. (2011). Discrete and continuous time representations and mathematical models for large production scheduling problems: A case study from the pharmaceutical industry. European Journal of Operational Research, 215, 383-392. Tang, L., & Jiang, S. (2009). The Charge Batching Planning Problem in Steelmaking Process Using Lagrangian Relaxation Algorithm. Industrial & Engineering Chemistry Research, 48, 7780-7787. Timpe, C. H., & Kallrath, J. (2000). Optimal planning in large multi-site production networks. European Journal of Operational Research, 126, 422-435. Tosserams, S., Etman, L. F. P., Papalambros, P. Y., & Rooda, J. E. (2006). An Augmented Lagrangian Relaxation for Analytical Target Cascading Using the Alternating Direction Method of Multipliers. Structural and Multidisciplinary Optimization, 31, 176-189. Tosserams, S., Etman, L. F. P., & Rooda, J. E. (2008). Augmented Lagrangian coordination for distributed optimal design in MDO. Int. J. Numer. Meth. Engng., 73, 1885-1910. Varma, V. A., Reklaitis, G. V., Blau, G. E., & Pekny, J. F. (2007). Enterprise-wide modeling & optimization--An overview of emerging research challenges and opportunities. Computers & Chemical Engineering, 31, 692-711. Varvarezos, D. Recent advances in planning and blending operations. In Refinery Optimization: Aspen Technology, Inc. 301 Velez, S., Sundaramoorthy, A., & Maravelias, C. T. (2013). Valid Inequalities Based on Demand Propagation for Chemical Production Scheduling MIP Models. AIChE Journal, 59, 872887. Verderame, P. M., & Floudas, C. A. (2008). Integrated Operational Planning and Medium-Term Scheduling for Large-Scale Industrial Batch Plants. Industrial & Engineering Chemistry Research, 47, 4845-4860. Verderame, P. M., & Floudas, C. A. (2009). Operational planning framework for multisite production and distribution networks. Computers & Chemical Engineering, 33, 10361050. Wassick, J. M. (2009). Enterprise-wide optimization in an integrated chemical complex. Computers & Chemical Engineering, 33, 1950-1963. Wilkinson, S. J., Cortier, A., Shah, N., & Pantelides, C. C. (1996). Integrated production and distribution scheduling on a Europe-wide basis. Computers & Chemical Engineering, 20, S1275-S1280. Wu, D., & Ierapetritou, M. (2007). Hierarchical approach for production planning and scheduling under uncertainty. Chemical Engineering and Processing: Process Intensification, 46, 1129-1140. Wu, D., & Ierapetritou, M. G. (2003). Decomposition approaches for the efficient solution of short-term scheduling problems. Computers & Chemical Engineering, 27, 12611276. Zhao, X., Liu, W., & Wang, J. (2008). Genetic Algorithm optimizer for blend planning. In Control Conference, 2008. CCC 2008. 27th Chinese (pp. 86-88).
© Copyright 2024